NETWORKED TRUST: COMPUTATIONAL
UNDERSTANDING OF INTERPERSONAL TRUST
ONLINE
A Dissertation
Presented to the Faculty of the Graduate School
of Cornell University
in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy
by
Xiao Ma
August 2019
© 2019 Xiao Ma
ALL RIGHTS RESERVED
NETWORKED TRUST: COMPUTATIONAL UNDERSTANDING OF
INTERPERSONAL TRUST ONLINE
Xiao Ma, Ph.D.
Cornell University 2019
My doctoral research develops a deeper understanding of interpersonal trust
online through computational methods, in the context of online exchange plat-
forms including peer-to-peer marketplaces, sharing economy platforms, and
social networks. Through analyzing images in product listings on eBay and
LetGo.com, language in profiles on Airbnb, and networks in social groups on
Facebook, I show how different algorithms help understand and predict inter-
personal trust in each context. Findings reveal patterns of interpersonal trust.
For example, high-quality images are perceived as more trustworthy than stock
imagery; language of promises lead to higher perceived trustworthiness through
conventional signaling; and smaller, denser, and more private social groups are
trusted more. These findings inform the design of online exchange platforms.
The algorithms predicting trust could also be used for better ranking and rec-
ommendation to “engineer“ interpersonal trust. Going forward, I propose a
lens of “networked trust” to view interpersonal trust online, which has three
focuses: (1) cues in Computer-Mediated Communication; (2) embeddedness
in social networks; and (3) increasing mediation by algorithms. The networked
trust framework can be used to frame future trust research in other contexts,
such as misinformation. Finally, two research agenda were charted by this dis-
sertation — AI-Mediated Communication and AI-Mediated Exchange Theory,
which future work can develop on.
BIOGRAPHICAL SKETCH
Xiao Ma is a PhD candidate in Information Science and a member of the
Social Technologies Lab at Cornell Tech, Cornell University. She started the
PhD in September 2014 at Cornell Tech’s temporary campus at the Google NYC
building in Chelsea before it moved to Roosevelt Island in July 2017. She is the
first PhD to graduate fully contained at the new campus of Cornell Tech since its
establishment in 2012.
During her PhD, Xiao has worked as a User Experience (UX) researcher at
Airbnb and Facebook, and as a data scientist/engineer at the Core Data Science
team at Facebook. Different roles have given her different perspectives on the
design, implementation, and analysis of real-word socio-technical systems. She
also holds a Bachelor of Science in Microelectronics from Peking University.
iii
To my parents, for giving me the courage and freedom to see a world of infinite
possibilities.
iv
ACKNOWLEDGEMENTS
The past five years at Cornell Tech have been such an incredible journey. I
had a front-row seat to witness the founding and the growth of Cornell’s campus
in New York City, Cornell Tech. I would like to thank everyone who made
Cornell Tech possible. It was an honor to see the amount of work that goes into
creating something new from scratch, and experience the change as we grow.
Entrepreneurship is a large part of my identity in addition to research, and I
hope to get a chance to build something from scratch of my own one day.
I believe the best part of the PhD is the people, and I have so many to thank.
First and foremost, my deepest gratitude goes to my advisor, Mor Naaman,
who not only taught me how to do good research, but also how to be a better
human being. Thank you for always seeing my potential and encouraging me
to achieve it. One of the most important lessons I learned from Mor is that
the questions matter equally if not more than the answers. I love the process of
brainstorming and discussing with Mor to formulate a question, articulate it, then
finally solve it. I also thank Mor for sharing his travel stories and introducing me
to more music. I probably didn’t travel as much as Mor did during his PhD, but
I definitely have been to more art exhibits.
I am also so fortunate to have a very interdisciplinary and supportive com-
mittee, including Jeff Hancock, Karen Levy, and Serge Belongie. Thank you
for teaching me on different fields, including communication, sociology, and
computer science. I’d also like to thank Natalie Bazarova, Clarence Lee, Louise
Barkhuus and Raz Schwartz for providing important feedback on my early work
during the PhD.
Outside of Cornell, I have also learned a lot from having support from other
institutions, including Stanford, Facebook, Airbnb, and Women in Technology
v
and Entrepreneurship in New York (WiTNY).
I visited the Stanford Institute for Research in the Social Sciences in my second
year of PhD. Thank you to Paolo Parigi and Bruno Abrahao for hosting me, and
Michael Bernstein for having me spend time with the Stanford HCI group. I
made good friends there including Ali Alkhatib, Niloufar Salehi, Ranjay Krishna,
and many others. I also reconnected with Scott Cheng, who was doing his
Master’s in HCI at Stanford at the time and later became my partner when he
moved to New York. Scott and I have very different personalities. His calmness
is always there for me whenever I experience anxieties from research.
Facebook and Airbnb opened the door for industry for me. I am grateful for
having experienced the unique cultures at both places. Thank you to Judd Antin,
Janna Bray, Yoni Karpfen, and many others at Airbnb for showing me the ropes
of industry research. Thank you to Jenna Lee, Tracy Mehlman, and Affonso Reis
for helping me think more deeply about career trajectory. Finally, thank you to
the Facebook Core Data Science team, especially Justin Cheng, Lada Adamic,
Shankar Iyer, Alex Dow, Moira Burke, Bogdan State, Carlos Diuk, for one of my
favorite summers in California. I have always admired Justin’s work and it was
an honor to be able to work together. One of my favorite memories during that
summer is the Computational Social Science group’s research writing retreat
Lada organized, when we all discussed research ideas around bonfire in the
mountains of Santa Cruz.
I was also so fortunate to be a Women in Technology and Entrepreneurship
in New York (WiTNY) fellow during my PhD, advised by Judy Spitz. Through
working with Judy, I developed more leadership skills and made efforts to
contribute back to the community. Judy is charismatic, resourceful. She is an
amazing leader, mentor, and role-model.
vi
I also want to thank all my lab mates and friends I made during graduate
school, including Emily Sun, Ross McLachlan, Nir Grinberg, Maurice Jakesch,
Kimberly Wilber, Hani Altwaijry, Andreas Veit, Neta Tamir, Justine Zhang, and
many others. I thank Amy X. Zhang at MIT, who convinced me to come to
Cornell Tech to work with Mor, and who has always been a role-model and
inspiration.
Finally, I would like to thank my parents, Jianping Ma and Yan Liang. My
father is an academic himself and I grew up on the campus of his university,
spending more time in the libraries and classrooms than on the playground.
He inspired me to be intellectually curious, and to have broad interests (he is a
physicist with interest in calligraphy, poetry and literature). My mother studied
computer science, and worked as a software engineer in the 1970s in China. She
raised me to be independent, strong, and taught me to write my first program.
My parents instilled in me by example the best ethics, in life, and about work.
They also gave me the freedom to travel, to see the world, and most importantly,
to explore all the possibilities and make my own decisions — they even gave
me a name that means freedom in Chinese. The freedom is especially valuable
coming from Chinese parents of their generation because of the single child
policy. Confucius says, "Don’t travel far away while your parents are still alive.
If you do, make it worthwhile." I hope I have made the PhD journey worthwhile,
and I hope my parents are proud and may they forgive me for all the time I am
absent away from home.
vii
TABLE OF CONTENTS
Biographical Sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1 Introduction 1
1.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Summary of Contribution . . . . . . . . . . . . . . . . . . . . . . . 12
2 Background 15
2.1 Theoretical Foundations of Trust . . . . . . . . . . . . . . . . . . . 16
2.2 Measuring Trust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Determinants and Outcomes of Trust — A Multilevel Per-
spective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Methodological Note on Measuring Trust . . . . . . . . . . 32
2.3 Online Trust and Networked Trust . . . . . . . . . . . . . . . . . . 37
2.3.1 Online Trust . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.2 Networked Trust . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Bias and Distrust . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3 Images of Trust: Understanding Image Quality and Trust in Peer-to-
Peer Marketplaces 48
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.1 LetGo.com . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.2 eBay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Annotating Image Quality . . . . . . . . . . . . . . . . . . . . . . . 56
3.5 Modeling Image Quality . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5.1 What Makes a Good Product Photo? . . . . . . . . . . . . . 60
3.6 Marketplace Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6.1 Image Quality and Sales . . . . . . . . . . . . . . . . . . . . 68
3.6.2 Perceived Trustworthiness . . . . . . . . . . . . . . . . . . . 70
3.7 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . 73
4 Language of Trust: Self-Presentation and Perceived Trustworthiness of
Airbnb Host Profiles 77
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
viii
4.3 Step 1: Self-Presentation and Perceived Trustworthiness of Airbnb
Host Profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3.1 Study 1 — How do Hosts Self-Disclose? . . . . . . . . . . . 85
4.3.2 Study 2 — Self-Disclosure and Perceived Trustworthiness 96
4.3.3 Study 3 — From Perception to Choice . . . . . . . . . . . . 106
4.4 Step 2: A Computational Approach to Perceived Trustworthiness
of Airbnb Host Profiles . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.4.1 Method and Dataset . . . . . . . . . . . . . . . . . . . . . . 113
4.4.2 Predicting Perceived Trustworthiness . . . . . . . . . . . . 115
4.4.3 Factors Contributing to Perceived Trustworthiness . . . . 120
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.5.1 Design Implications . . . . . . . . . . . . . . . . . . . . . . 125
4.5.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.6 Conclusion and Extensions . . . . . . . . . . . . . . . . . . . . . . 127
5 Networks of Trust: When Do People Trust Their Social Groups? 131
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.2.1 Determinants of Trust in Groups . . . . . . . . . . . . . . . 135
5.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.3.1 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.3.2 Survey Design . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.3.3 Data and Statistical Approaches . . . . . . . . . . . . . . . 140
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.4.1 Individual Differences and Trust . . . . . . . . . . . . . . . 143
5.4.2 Group Differences and Trust . . . . . . . . . . . . . . . . . 144
5.4.3 Predicting Trust in Groups . . . . . . . . . . . . . . . . . . 153
5.4.4 Group Outcomes . . . . . . . . . . . . . . . . . . . . . . . . 154
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.5.1 Design Implications . . . . . . . . . . . . . . . . . . . . . . 158
5.5.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6 Discussion and Future Work 162
6.1 Algorithms of Trust and Networked Trust . . . . . . . . . . . . . . 162
6.2 Future Research Agenda . . . . . . . . . . . . . . . . . . . . . . . . 164
6.2.1 Networked Trust and Misinformation . . . . . . . . . . . . 164
6.2.2 AI-Mediated Communication (AI-MC) . . . . . . . . . . . 168
6.2.3 AI-Mediated Exchange Theory (AI-MET) . . . . . . . . . . 170
6.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Bibliography 177
ix
LIST OF TABLES
3.1 Image feature definitions and example images . . . . . . . . . . . 63
3.2 Ordered logistic regression coefficients predicting image quality
labels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3 Image quality predicted by our models is positively associated
with higher likelihood that an item is sold (1.17x more for shoes,
and 1.25x more for handbags). . . . . . . . . . . . . . . . . . . . . 69
3.4 Marketplace perceived trustworthiness scale . . . . . . . . . . . . 74
4.1 Topics of self-disclosure in Airbnb host profiles. . . . . . . . . . . 87
4.2 Six-item perceived trustworthiness scale. . . . . . . . . . . . . . . 98
4.3 Performance of sentence category classification. . . . . . . . . . . 118
4.4 Model performance (accuracy) summary by different length batch.
Random baseline accuracy is 50%. Models compare profiles of
similar lengths to predict relative perceived trustworthiness based
on word count. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.5 Factors contributing to higher and lower perceived trustworthi-
ness by different length batch. LIWC categories “we” and “social”
are important positive indicators, potentially through social war-
ranting. The LIWC “sexual” category contained mostly the word
“love”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.1 Trust in groups survey items. Participants reported the degree to
which they agreed or disagreed to each of the survey items on a
five-point Likert scale. . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.2 Descriptive summary of survey measures, including general atti-
tudes and trust in groups. Sparklines represent the histogram of
each measure. (N=6,383) . . . . . . . . . . . . . . . . . . . . . . . 142
5.3 Baseline model predicting trust in groups using demographics,
disposition to trust, risk attitude, in-group loyalty, and social
support. (N=6,323 after removing missing age and gender obser-
vations) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.4 Five sets of group-level features used for predicting trust in groups.145
x
LIST OF FIGURES
2.1 Direct and two forms of generalized (indirect) exchange illus-
trated by Bearman 1997. Each character represents a party in-
volved in the exchange. In direct social exchange, there is direct
transfer of value or goods between A and B. In generalized ex-
change, multiple parties pool their resources together to produce
greater value, which is then further distributed. . . . . . . . . . . 19
2.2 Differences among trust, trustworthiness and perceived trustwor-
thiness. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Multilevel perspective of trust. . . . . . . . . . . . . . . . . . . . . 24
2.4 The integrative organizational trust model proposed by Mayer et
al. 1995 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1 We study the interplay between image quality, marketplace out-
comes, and user trust in peer-to-peer online marketplaces. Here,
we show how image quality (as measured by a deep-learned CNN
model) correlates with user trust. User studies in Sec. 3.6.2 show
that high quality images selected by our model out-performs
stock-imagery in eliciting user trust. . . . . . . . . . . . . . . . . 51
3.2 Samples of lowest-rated and highest-rated images from shoes
and handbags groundtruth. . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Left: standardized score distributions for filtered images on shoes
and handbags categories. Right: final ground truth labels, show-
ing a fairly even distribution. . . . . . . . . . . . . . . . . . . . . 57
3.4 Hypothetical marketplace mock-ups used for our user experi-
ment. From left to right, showing images with high quality score,
low quality score, and stock imagery. . . . . . . . . . . . . . . . . 70
4.1 Probability of disclosure per topic . . . . . . . . . . . . . . . . . . 94
4.2 Average number of sentences per topic . . . . . . . . . . . . . . . 94
4.3 Self-disclosure trends by topic and host type. The error bars
represent one standard error. . . . . . . . . . . . . . . . . . . . . . 94
4.4 Perceived trustworthiness increases with profile length (x-axis on
log scale). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5 Perceived trustworthiness score distributions for profiles with
different number of topics. . . . . . . . . . . . . . . . . . . . . . . 103
4.6 Comparison for different strategies, organized by the number
of topics mentioned in the profile. The dotted lines indicate the
bottom and top quartiles in each topic-count group. . . . . . . . . 104
4.7 Observed likelihood that the host with high trustworthiness score
is preferred and the probability predicted by Bradley-Terry model,
by profile length. The error bars represent one standard error. . . 110
xi
5.1 The relationship between trust in groups and group size, for
each dimension (panels), across groups of different privacy types
(line style) and individuals with different propensity to trust (line
color). Dunbar’s number (150) is marked by a vertical red dotted
line.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.2 People have the highest trust in friends and family groups, and
lowest in interest- and location-based groups. . . . . . . . . . . . 149
5.3 Groups differ in network density, participant degree centrality,
and how a participant’s friends are linked to each other. Each
node represents a group member. Each edge represents a friend-
ship between two members. The survey participant is colored in
red, and group admins are colored in yellow. . . . . . . . . . . . . 152
5.4 For each feature set, we calculated the average feature importance
(measured by relative percent increase in MSE when a feature is
removed) in predicting trust in groups. Network structure was
the most important, followed by an individual’s general attitudes
towards others. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.5 Groups with higher trust ratings are less likely to increase in
size (left), more likely for the survey participant to form new
connections in them (right), and had no effect on the likelihood
on forming friendships among group members other than the
rater (center). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
xii
CHAPTER 1
INTRODUCTION
Trust, in the broadest sense of confidence in
one’s expectations, is a basic fact of social life.
Niklas Luhmann
in Trust and Power (1979)
Our world builds on the foundation of social exchange. When we say social
exchange, we refer to both the psychological and economic aspects of social
relationships [Emerson, 1976]. According to Homans, elementary forms of social
behavior such as communication among group members can be viewed as an
exchange [Homans, 1958]. Employment relationships can also be viewed as an
exchange where the reward to the employee is the wage, and the cost is the
responsibilities they have [Homans, 1958]. Viewing social relations as exchanges,
especially as exchanges in social networks, is helpful in understanding phenom-
ena in social structures, including group cohesiveness [Homans, 1958], social
solidarity [Bearman, 1997], and power [Blau, 1964].
One of the most important concept that enables the functioning of social
exchange systems at scale is trust. Because there are often risks involved in a
potentially mutually beneficial exchange, trust is needed to overcome the risks
and to facilitate the exchange. For example, commerce could potentially be
beneficial for both the buyer and the seller. But the buyer must take “a leap of
faith” and trust the seller to provide accurate information about the merchandise
before making the decision to purchase. Without trust, no exchange could take
place under risks and hardly any form of collaboration could exist. For each
1
individual, without trust, the everyday social life which we take for granted
is simply not possible. As Niklas Luhmann wrote in opening chapter of his
foundational book Trust and Power [Luhmann, 1979]:
“[...] a complete absence of trust would prevent him or her [a person]
from even getting up in the morning. [...] Such abrupt confrontation
with the complexity of the world at its most extreme is beyond human
endurance.”
As trust mitigates the uncertainty and complexity in social exchange systems,
trust has been frequently attributed for systems’ success. A deeper understand-
ing of trust is important for the social, economic and political outcomes of
social systems — including for individuals, groups, and nations. For example,
trust is considered as an important variable in the development of healthy fam-
ily relationships and personalities for individuals [Rotter, 1967] Trust has also
been shown to be important for the productivity and cohesiveness of organiza-
tions [Mayer et al., 1995, Nyhan, 2000, McEvily et al., 2003, Fine and Holyfield,
1996]. Finally, Fukuyama showed that trust leads to the creation of prosperity for
nations [Fukuyama, 1995].
Because of the importance of trust, a robust line of research exists across
disciplines on the factors that lead to trust, and the outcomes that trust contribute
to, in sociology, psychology, economics, marketing, and management science.
At the same time, new questions about trust, especially trust in socio-technical
systems, emerge as a result of increasing digital mediation on social exchange
relationships. In the past few decades, social exchange relationships have under-
gone fundamental changes. In my view, the digitalization of social exchange
took place in three waves: (1) the digitization of the exchange of goods; (2) the
2
digitization of social relationships; and (3) the digitalization of the exchange of
resources through sharing economy platforms. Each of these waves brought
significant changes to social exchange structures, raising new questions about
trust at the same time. The work presented in this dissertation addresses these
new questions through the development of computational methods.
What are some of the new questions about trust that each wave of the digital-
ization of social exchange brought?
The first wave, the digitization of the exchange of goods, was enabled by
online commerce platforms including Amazon (founded in 1994) and eBay
(founded in 1995). Through these digital exchange platforms, sellers can reach
customers widely regardless of location, and customers in turn have much wider
selections to choose from compared to their local stores. However, one core risk
emerged — if the buyers pay first, it is possible that the sellers would not send
the goods (and disappear from the Internet and reappear under a different name),
or send defect goods that do not meet the expectations of the buyers; however, if
the sellers send the goods first, the buyers could refuse to pay. In other words,
people have to trust others with their financial resources for exchanges to take
place on these online commerce platforms. Multiple mechanisms emerged to
mitigate the risks and to increase trust, including the development of third-party
online payment systems that acted as an assurance (e.g., PayPal, founded in
1998). Reputation systems also appeared as a key mechanism to increase trust.
A long line of research on reputation systems has discussed their benefits and
limitations [Resnick et al., 2000, Resnick and Zeckhauser, 2002, Resnick et al.,
2006, Cook et al., 2009].
Two decades later, new questions about trust on these online commerce
3
platforms are still emerging as platforms continue to evolve. For example, due to
the wide accessibility of mobile photography, user-generated images of goods
play an ever increasingly important role in establishing trust in peer-to-peer
marketplaces, compared to stock imagery. New advancements in computer
vision (e.g., deep learning) also present opportunities to understand how user-
generated images establish trust. This dissertation presents a case study on
leveraging new computational opportunities to understand and predict how
user-generated images impact trust in online marketplaces.
The second wave of the digitalization of exchange was the digitization of
social relationships, characterized by the ubiquitous adoption of social network
sites (SNSs) such as Facebook (founded in 2004), Twitter (founded in 2006),
LinkedIn (founded in 2002) and Instagram (founded in 2010). Although online
communities have long existed since the beginning of the Internet, these SNSs
enable users to articulate and make visible their social networks [Boyd and Elli-
son, 2007]. New questions around trust again emerged. The core risk presented
by SNSs in social exchange is how personal information flows through networks.
Self-disclosure on SNSs can be visible to a “networked audience” [Marwick
and Boyd, 2011], including real and potential viewers through a broadcasting
audience network. Therefore, there are a lot of risks when people post or share
content online. People need to trust others with their personal information, includ-
ing information about their identities, biometric data, social ties, experiences,
feelings, emotions, etc.
Important opportunities are also brought by SNSs to better understand in-
terpersonal trust, especially around how tie strength and network structure
contribute to interpersonal outcomes. For example, computational methods
4
have been developed to predict strong and weak ties [Gilbert and Karahalios,
2009]. People’s friendship network structure has been found to be informative
for predicting romantic relationships [Backstrom and Kleinberg, 2014]. This
dissertation continues to leverage the opportunities to study network structure
by presenting a large-scale analysis on how social network structure predicts
trust in Facebook groups.
Finally, the third wave of the digitalization of exchange — the digitalization
of the exchange of resources — took place at scale through sharing economy
platforms such as Airbnb (founded in 2008); Uber (founded in 2009); and Lyft
(founded in 2012). Building on top of the payment and social networks infras-
tructure developed by the first two waves of the digitalization, the last wave of
the digitalization of social exchange requires that people not only trust others
with their financial resources, personal information, but also their physical safety.
When people stay at strangers’ homes (Airbnb), or hop into strangers’ cars (Uber
and Lyft), a great amount of risks are being mitigated through interpersonal
trust as well as trust in the sharing economy platforms. One of the most distinct
characteristics of these platforms is that they blur the line of social and econom-
ical exchange [Hamari et al., 2016, Ikkala and Lampinen, 2015, Lampinen and
Cheshire, 2016], For example, in the context of Airbnb, “guests” stay at “hosts”’
homes paying a fee, constituting an economic transaction. However, unlike a
purely hotel-like interaction, some Airbnb hosts also value social interactions
as well as potential friendships built through hosting [Lampinen and Cheshire,
2016].
The in-between nature of sharing economy interactions creates new questions
about how people trust each other on these platforms. Social mechanisms such
5
as homophily can play a role in establishing trust [Abrahao et al., 2017]. The
problem of bias can also arise [Edelman and Luca, 2014]. At the same time,
reputation systems can help offset homophily tendencies [Abrahao et al., 2017].
Other mechanisms such as self-disclosure can also establish trust by reducing
uncertainty [Berger and Calabrese, 1975, Ellison and Hancock, 2013]. This disser-
tation presents a case study on self-disclosure and trust through computational
analysis of the language in host profiles on Airbnb. As sharing economy con-
tinues to evolve and mature, new questions about trust will continue to arise.
For example, in addition to lodging, Airbnb has also launched “experiences”1,
where hosts lead in-person tours for guests. “Experiences” require higher levels
of interpersonal interaction and hence also higher levels of trust. Understanding
how trust functions at different levels and how to predict trust computationally
by level remain future work.
Taken together, three waves of the digitalization of exchange have fund-
mentally altered the social structure under which people conduct exchanges
— creating new questions about how people trust each other within these new
structures. My work address some of these new questions about trust in the
context of each wave, with a focus on leveraging computational methods to build
algorithms to understand and predict trust. Through computational methods
on analyzing images, language, and social networks, we gain a deeper under-
standing of trust in digitalized social exchange that are perhaps hard to obtain
with qualitative or smaller scale studies. The insights on how people trust each
other in each context can inform the design of the digital platforms facilitating
exchange. At the same time, these new computational models also have the
potential to be built directly into the digital platforms to influence the ranking
1https://www.airbnb.com/help/article/1581/what-are-experiences
6
and recommendation of potential exchange partners, fostering trust through
engineering.
More importantly, the work presented in this dissertation prompts us to think
about trust in digitalized exchange differently. Originally, trust as a concept
was developed as a social construct between people [Luhmann, 1979]. The concept
of online trust, or eTrust, gained importance with the first wave of digitalization of
exchange in the context of trust on online commerce platforms [Cook et al., 2009].
However, as discussed in the contexts of SNSs, trust is also affected by networks.
In addition, increasingly, algorithms are gaining more control over how we trust
online through personalization and recommendations. Online trust as a term or
framework for thinking about trust in these socio-technical systems is becoming
insufficient.
To incorporate and emphasize the roles of networks and algorithms in
trust in digital platforms that facilitate exchange, I propose the framework of
“networked trust”. The term “networked” is borrowed from the concept of “net-
worked individualism”, which refers to the new social structure where people
are embedded in social networks with more weak ties as a result of digitalization
of social relationships [Rainie and Wellman, 2012]. There are three focuses in net-
worked trust: (1) cues in Computer-Mediated Communication; (2) embeddedness
in social networks; and (3) increasing mediation by algorithms.
The first focus of networked trust, cues in Computer-Mediated Communi-
cation (CMC), refers to how cues that are available in digitalized social exchange
differ from cues in face-to-face interactions. This focus of networked trust is
most aligned with the traditional meaning of the term “online trust”. Increased
uncertainty brought by CMC results in the need for overcoming such uncertainty
7
to establish trust. In my work, I discuss how images and language cues play a
role in establishing trust in Chapter 3 and Chapter 4, by leveraging opportunities
brought by new computational methods to deepen our understanding of cues.
The second focus of networked trust, embeddedness in social networks,
refers to how information about social networks can affect trust. For example, on
LinkedIn, when users view another person’s profile, LinkedIn informs the users
about how they are connected in the LinkedIn network. Information such as the
number of common connections, the degree of separation, the specific people
that can act as a bridge to introduce the users, are important for establishing
trust in the context of networking. In my work, Chapter 5 shows how networks
contribute trust in the context of social groups on Facebook.
Finally, the third focus of networked trust, increasing mediation by algo-
rithms, refers to how trust can be impacted by algorithmic reconfiguration of
social exchange relationships. It is no secret that modern digitalized exchange
platforms use algorithms for personalization [Gubin et al., 2017, Bakshy et al.,
2015, Grbovic and Cheng, 2018, Linden et al., 2003, Zhou et al., 2008]. Notable
examples of such algorithmic mediation include, Facebook News Feed rank-
ing [Gubin et al., 2017, Bakshy et al., 2015], Netflix Prize [Zhou et al., 2008], and
Amazon’s item-to-item collaborative filtering [Linden et al., 2003]. Increasingly,
sharing economy platforms such as Airbnb also deploy algorithmic ranking and
matching algorithms to recommend exchange partners, potentially optimizing
for conversion [Grbovic and Cheng, 2018]. A key part of the utility value of
Uber and Lyft comes from the algorithmic matching of drivers and passengers
through dispatch and routing. These algorithms create efficiency in exchange
relationships, but can also create problems for trust. The opaqueness and invisi-
8
bility of such algorithms can hinder not only how users trust the system [Eslami
et al., 2015], but also how people trust each other when mediated by algorithms.
The algorithmic mediation in interpersonal exchange also raises new questions
about bias, which can further reduce trust in the platform. For example, should
matching algorithms on online dating platforms recommend potential partners
for users of their own race [Hutson et al., 2018]? This dissertation does not di-
rectly address this focus of networked trust, but Chapter 6 discusses how future
work can make an important theoretical contribution through the development
of AI-Mediated Exchange Theory.
The work presented in this dissertation sets the foundation for several areas
of future work. The networked trust lens can be directly applied to new contexts
to understand how different factors affect trust in a networked environment. For
example, in the context of misinformation, networked trust framework will look
at how factors work together to impact trust in information, including cues (the
credibility of news sources), networks (how information spreads through social
networks), and algorithms (the impact of algorithmic curation).
At the same time, networked trust can be extended to chart two future direc-
tions of research: AI-Mediated Communication (AI-MC) and AI-Mediated Ex-
change Theory (AI-MET). The first focus of networked trust, cues in Computer-
Mediated Communication, can be extended to AI-Mediated Communication
(AI-MC). In interpersonal communication, increasingly AI-powered systems can
modify, augment, and even generate cues in self-presentation or messages to
optimize for interpersonal outcomes. For example, the algorithms built in this
dissertation predicting trust in Airbnb host profiles can be fitted to improve a
given profile to make it appear to be more trustworthy. AI-MC aims to under-
9
stand how such AI-powered systems might impact interpersonal communication
outcomes, which is important for the future of networked trust. Early studies
have already observed initial effects of AI-MC [Jakesch et al., 2019] while exact
definition and research agenda are being developed and completed [Naaman
et al., 2019].
At the same time, the third focus of networked trust, increasing mediation
by algorithms, can be extended to AI-Mediated Exchange Theory (AI-MET).
AI-MET is an extension of social exchange theory [Cheshire, 2007,Yamagishi and
Cook, 1993,Bearman, 1997]. By viewing AI systems as mediating social exchange
in socio-technical systems, AI-MET articulates different mechanisms through
which AI reconfigures social exchange relationships. AI-MET can allow different
perspectives in exploring the relationships between humans and AI to speak to
each other. AI-MET can be useful in understanding issues of AI in socio-technical
systems such as how trust in the algorithms affects trust in another person when
their relationships are mediated by algorithms.
1.1 Outline
The rest of this dissertation is organized as such: Chapter 2 first sets a theoretical
foundation for the definition of trust using social exchange theory. Chapter 3
- Chapter 5 each details a case study on the computational understanding and
prediction of trust. Each case study investigate trust in the context of a platform
representing specific wave of the digitalization of social exchange, including
peer-to-peer marketplaces, social networks, and sharing economy platforms.
Then, Chapter 6 charts future research agenda based on the work presented in
10
this dissertation.
Specifically, Chapter 2 reviews the literature that my work builds upon. The
first section reviews social exchange theory as the theoretical foundation for
the definition of trust. Through social exchange theory, we view the world as
an equilibrium established through negotiated exchange. Trust mitigate risks
during exchange, leading to higher efficiency and better societal outcomes. Then,
key literature on the determinants, outcomes, and the measurement of trust
was reviewed through a multilevel perspective. Special attention was paid to
literature in online trust, and the lens of networked trust was discussed in more
detail in relation to literature in online trust. Finally, literature related to the
flip side of trust, bias was reviewed, including the racial, gender, status, and
algorithmic bias.
In the two chapters that follow, I cover work on the first focus of networked
trust — how trust is established through images (Chapter 3) and language
(Chapter 4) cues.
Chapter 3 examines the role of image cues on trust in the context of online
peer-to-peer buy-and-sell marketplaces. Using techniques from computer vision,
including both image feature-based and deep learning, I show that we can
predict image quality with high accuracy (close to 90%), and that high quality
images predicted using our model outperform stock imagery in generating trust
in marketplaces. Finally, I also show that higher quality images predicted by our
model leads to higher sales, showing how algorithms predicting trust translate
to real-world outcomes.
Chapter 4 discusses how we can learn about trust on Airbnb based on lan-
11
guage cues in host profiles. There are two parts in this chapter. First, I discuss
how we constructed the dataset of Airbnb host profiles with perceived trust-
worthiness, while developing different topics that are frequently mentioned in
the profiles. Then, I present a computational framework to predict trust in host
profiles leveraging natural language processing and machine learning.
Next, Chapter 5 expands on the second focus of networked trust — trust
in social exchange that is embedded in social networks. Chapter 5 presents a
large-scale comprehensive study of trust in people’s social groups on Facebook.
Leveraging survey data with social network and interactions data, we learn the
patterns of trust across a diverse range of groups of different sizes and categories.
I show how both individual characteristics and group characteristics such as
group network size and density matter for trust in groups. In addition, I show
how trust in groups may lead to different group outcomes, and how we can
leverage these findings to design for better online groups.
Finally, Chapter 6 charts three directions for future work: (1) networked
trust and misinformation; (2) AI-Mediated Communication (AI-MC); and (3)
AI-Mediated Exchange Theory (AI-MET).
1.2 Summary of Contribution
This dissertation contributes to the field of Human Computer Interaction (HCI),
Computer-Supported Cooperative Work and Social Computing (CSCW), and
Computational Social Science in the following ways:
1. Through three comprehensive studies of computational trust in different
12
contexts, this dissertation extends previous understanding of trust lever-
aging new techniques (e.g., deep learning), new contexts (e.g., sharing
economy), and new data sources (e.g., network structure). Each study
showed that not only we were able to understand and predict trust using
a variety of different algorithms, but also trust has real-world impact on
specific outcomes. Importantly, all work presented in this dissertation is
based on large-scale real-world systems — covering major categories of
social exchange platforms that represent different waves of digitalization
of social exchange (online commerce, social networks, and sharing econ-
omy). The findings in this dissertation validate and extend what we know
about trust, and can inform the design of future digital platforms to better
facilitate trust.
2. This dissertation proposes a framework of “networked trust”, which can
be a useful lens to further the discussion from “online trust” to focus more
on factors that are increasingly important — networks and algorithms.
Networked trust can be applied to different contexts where trust is urgently
needed, such as misinformation, to understand how cues, networks, and
algorithms work together to affect trust.
3. Finally, networked trust can further be extended to chart two important
research agenda for future development — AI-Mediated Communication
and AI-Mediated Exchange Theory. AI-Mediated Communication focuses
on extending the cues focus of networked trust, by better understanding
the impact of AI on interpersonal outcomes when cues are subject to the
modification, augmentation, and generation by AI. AI-Mediated Exchange
Theory, on the other hand, focuses on incorporating the role of algorithms
in networked social exchange broadly. Early work has been established
13
in AI-Mediated Communication [Jakesch et al., 2019, Naaman et al., 2019],
while AI-Mediated Exchange Theory remains future work.
14
CHAPTER 2
BACKGROUND
Trust is the chicken soup of social life. [...] Like
chicken soup, trust appears to work somewhat
mysteriously.
Eric M. Uslaner
in Producing and Consuming Trust (2000)
The work presented in this dissertation is highly inter-disciplinary. My defi-
nition and the scoping of trust build on previous scholarship across disciplines,
just to name a few, sociology [Luhmann, 1979, Gambetta, 1988], psychology [Rot-
ter, 1967], economics [Berg et al., 1995], business [Grabner-Kraeuter, 2002], and
political science [Miller, 1974]. This chapter reviews prior scholarship on trust
across disciplines that my work builds on through a multilevel perspective.
One key challenge of researching trust is that the term is heavily loaded
with many different meanings and bears a great deal of ambiguity. To clarify
the concept being studied, this chapter first provides a theoretical framework
under which we define trust, heavily drawing from social exchange theory (Sec-
tion 2.1). Then, in order to talk about trust concretely in specific contexts, I review
various measurements, determinants and outcomes of trust in prior literature
(Section 2.2). The work presented in this dissertation draws from the defini-
tion and measurements reviewed here. In addition, because this dissertation
focuses on trust in digitalized exchange, I devote a section to specifically focus
on previous literature on online trust (Section 2.3). As networks and algorithms
play a more important role in digitalized exchange, I propose the framework of
15
“networked trust” as a lens to reason about trust in networked environments and
to address new challenges that arise.
Finally, trust as a mechanism to facilitate exchange has key limitations. Trust-
ing someone usually means distrusting others. As the process of forming trust
can be subjective, bias can occur when there is need to trust. Section 2.4 reviews
literature on four different types of biases, racial, gender, status, and algorith-
mic. Bias can be manifested in algorithms in subtle ways. Before deploying
computational models that predict trust to real-world applications, we need to
be cautious and aware of the potential biases that these algorithms may have.
2.1 Theoretical Foundations of Trust
Trust is an inherently social phenomenon. To research trust and to connect
the results meaningfully to societal outcomes, it is important to view it not in
isolation but rather grounded in broad sociological traditions and structures.
This section provides the theoretical foundation for defining trust — drawing
from social exchange theory in sociology.
Sociology, the study of society, has multiple paradigms, which are “funda-
mental images of the subject matter within a science” [Ritzer, 1975]. Paradigms
are useful to “define what should be studied, what questions should be asked,
how they should be asked, and what rules should be followed in interpreting
the answer obtained” [Ritzer, 1975]. According to Ritzer, major paradigms in
sociology include: structural functionalism (or systems theory), conflict theory,
symbolic interactionism, and social exchange theory [Ritzer, 1975]. Structural
functionalism is oriented to analyze social structures and institutions, viewing
16
individuals as largely controlled by the structures they are embedded in. Conflict
theory, on the other hand, views the world in constant conflicts and focuses on
studying disintegration and change in society. Symbolic interactionism places
greater emphasize on the mental process of the individuals as active creators of
social reality. Finally, social exchange theory views social change and stability as
a process of negotiated exchanges between different parties, primarily driven by
the process of reinforcement [Emerson, 1976].1
We can view trust under any of these paradigms, and the resulted research
would have different focuses. For example, trust when viewed under structural
functionalism serves as a “mechanism to reduce social complexity” [Luhmann,
1979].2 Such a viewpoint can be found in one of the earliest classic work on
trust: the 1979 book Trust and Power by Niklas Luhmann [Luhmann, 1979]. The
complexity of social systems is due to both the scale as well as uncertainties
in interacting with each agent in the system. Analyzing the motivation and
potential behaviors of each agent is cognitively impossible. Trust provides a
shortcut for decision-making when navigating social systems and thus reduces
social complexity. As the “lubricant of social interaction” [Arrow, 1974], trust
facilitates social interactions, reducing the whole system’s complexity. On the
other hand, if we view trust under conflict theory, we can focus more on the
dynamics of distrust and mistrust [Kramer, 1999, Ely, 1980, Marsh and Dibben,
2005, Kiousis, 2001]. Trust under symbolic interactionism would include work
that focuses on the psychological process such as how nonverbal cues affect
trust in partners [Lee et al., 2013, DeSteno et al., 2012]. Finally, trust under
1We do not expand on the limitations of each paradigm in this dissertation as it is not the
focus, but curious readers can find relevant discussions in [Ritzer, 1975, Emerson, 1976].
2It is worth noting that trust is not the only mechanism to reduce social complexity. Power and
control can also reduce social complexity by reducing the freedom of others, through possibly
coercion, contracts, and reputation systems [Luhmann, 1979].
17
social exchange theory would focus on how trust enables cooperation — as
seen in another classic book on trust, Trust: Making and Breaking Cooperative
Relations [Gambetta, 1988].
In this dissertation, I draw primarily on the last paradigm, social exchange
theory. Viewing trust under social exchange theory has the following advantages:
(1) It allows for a clean definition of trust in the context of risks in exchange; (2) It
allows for a multilevel view of trust as social exchange theory accounts for both
micro and macro exchanges; (3) It allows for the easy incorporation of digital
platforms in exchange. We expand on each point below.
First, as mentioned above, trust is a loaded term, used colloquially to refer to
a variety of different things. Grounding trust in social exchange theory allows
for a clean definition of trust in the exchange context that mitigates risks. Social
exchange theory views social change and stability as a process of negotiated
exchanges between different parties. Social relations are the units of analy-
sis [Emerson, 1976] in social exchange theory. By analyzing costs and rewards
in these relations, social exchange theory reasons about how people make de-
cisions in the social world quasi-economically. In a potential exchange context,
risks can create obstacles of collaboration, preventing the exchange from taking
place. Trust, a decision to be vulnerable to the other party involved in the exchange, can
mitigate risks and thus enable exchange and cooperation.
As Diego Gambetta defined:
“Trust (or, symmetrically, distrust) is a particular level of the sub-
jective probability with which an agent assesses that another agent
or group of agents will perform a particular action, both before he
18
can monitor such action (or independently of his capacity ever to
be able to monitor it) and in a context in which it affects his own
action. When we say that the probability that he will perform an ac-
tion that is beneficial or at least not detrimental to us is high enough
for us to consider engaging in some form of cooperation with him.
Correspondingly, when we say that someone is untrustworthy, we
imply that the probability is low enough for us to refrain from doing
so.” [Gambetta, 1988]
Figure 2.1: Direct and two forms of generalized (indirect) exchange il-
lustrated by Bearman 1997. Each character represents a party
involved in the exchange. In direct social exchange, there is
direct transfer of value or goods between A and B. In general-
ized exchange, multiple parties pool their resources together to
produce greater value, which is then further distributed.
Second, unlike pure macro-focused theory such as structural functionalism,
social exchange theory focuses both on direct (micro) and generalized (macro)
exchange [Emerson, 1976]. Direct exchange usually refers to dyadic exchange,
19
where two parties engage in a mutually beneficial exchange relationship, trans-
ferring value or goods directly between two parties. At the same time, social
exchange theory also discusses generalized exchange — or sometimes referred to
as “productive exchange” — the process where resources from multiple parties
are pooled together to produce greater value, which is then further be dis-
tributed [Emerson, 1976, Bearman, 1997, Yamagishi and Cook, 1993]. Figure 2.1
contains illustrations of direct exchange (top left), and two different forms of
generalized exchange (top right: net form, bottom: chain form) [Bearman, 1997].
We can view trust in the context of facilitating both direct and generalized ex-
change — leading to a multilevel view of trust that we expand in Section 2.2.
The multilevel view of trust refers to the study of trust at different levels of
abstractions, for example, trust at the individual level, trust in groups, trust in
organizations, trust in the government, and most broadly, trust in the society.
A multilevel view of trust is useful for organizing prior work on trust across
multiple disciplines and traditions. At the same time, it can also inform the
development of computational models that incorporate predictors on both micro
and macro levels.
Finally, the exchange view of trust allows for easy incorporation of the ef-
fects of online platforms on social exchange and trust. The digital platforms
alter the social structure under which people conduct exchange, creating new
questions about trust during the process. In the past few decades, our social
exchange structure has undergone fundamental changes through three waves of
the digitalization of exchange: (1) the digitization of the exchange of goods; (2)
the digitization of social relationships; and (3) the digitalization of the exchange
of resources through sharing economy platforms. As a result, people rely on cues
that are available online such as images and language to establish trust, rather
20
than relying on nonverbal cues in face-to-face interactions. In addition, how
people are embedded in social networks plays a key role in establishing trust.
Defining trust under the framework of social exchange theory also allows for the
easy incorporation of networks as factors that impact trust in mediated exchange.
In fact, the “networked trust” framework proposed by this dissertation is such
an extension from social exchange theory [Cook and Whitmeyer, 1992].
In conclusion, the work presented in this dissertation views trust through the
lens of social exchange theory, one of the many sociological paradigms. Through-
out this dissertation, trust at the highest level of abstraction is defined as: a
willingness to make oneself vulnerable to other parties in direct or general-
ized exchange, which mitigates risks and enables exchange and cooperation.
However, in specific contexts, trust can be defined and measured in more context-
specific ways. We review literature on measuring trust next.
2.2 Measuring Trust
When I tell people that I conduct research on trust, the first question they usually
ask is, “Well, but how do you measure trust?” It is a fair question. The first step
to research anything is to define and measure it, especially for a loaded term such
as trust. This section focuses on the measurement of trust through first providing
a categorization of common definitions of trust, and then reviewing literature on
three key methods: (1) surveys; (2) real-world behavior proxies; and (3) lab-based
trust games. In particular, I use a multilevel perspective to categorize common
definitions of trust to provide some clarity on what trust means across different
disciplines and traditions. In addition, as I show in Chapter 5, a multilevel
21
perspective of trust can inform the development of computational models by
incorporating both micro and macro level factors to improve prediction accuracy.
Figure 2.2: Differences among trust, trustworthiness and perceived trust-
worthiness.
Before we delve in, it is helpful to first clarify a few closely related concepts:
trust, trustworthiness, and perceived trustworthiness. Hardin [Hardin, 2002] and
Kiyonari et al. [Kiyonari et al., 2006] pointed out that the general literature on
trust typically overlooked the difference between trust and trustworthiness, and
used the term interchangeably. As illustrated in Figure 2.2, trust is exhibited by a
trustor — a decision to be vulnerable, and trustworthiness is a characteristic of the
trustee [Kiyonari et al., 2006]. In particular, there are also important differences
between trustworthiness and perceived trustworthiness. Trustworthiness refers
to whether the trustee behaves in a way that is not harmful to the trustor in
reality even though that it is well within the trustee’s capacity and freedom
to do so [Kiyonari et al., 2006]. Perceived trustworthiness, on the other hand,
is in the eyes of the trustor. It refers to a subjective belief of the trustor about
the trustworthiness of the trustee, which can be of course be subjective and
biased [Cheshire, 2011].
The relationships among three concepts are complex. In one-shot interactions,
trust does not necessarily beget trustworthiness [Kiyonari et al., 2006]. Perceived
trustworthiness does not necessarily lead to trust either. The decision to trust
might be mediated by the risk tolerance of the trustor. In this dissertation, when
22
necessary, I try to distinguish between trust, trustworthiness, and perceived
trustworthiness. Such distinction is helpful for measuring of trust. But when
I discuss general concepts, such as in the context of “networked trust”, for
simplicity, I use “trust” to encapsulate trust, trustworthiness, and perceived
trustworthiness.
2.2.1 Determinants and Outcomes of Trust — A Multilevel Per-
spective
As a concept central to social exchange, trust has received research attention
from a variety of disciplines in the past, just to name a few, sociology [Luhmann,
1979, Gambetta, 1988], psychology [Rotter, 1967], economics [Berg et al., 1995],
business [Grabner-Kraeuter, 2002], and political science [Miller, 1974]. Different
traditions, focuses, and levels of abstractions across disciplines make it difficult
to reach an agreement on how trust should be measured, what leads to trust,
and its outcomes. To bridge the gaps among different disciplines, I provide a
multilevel perspective of prior literature on trust, focusing on the determinants
and outcomes of trust first, and its measurement in Section 2.2.2.
What is a multilevel view of trust? When people talk about trust, they can
be referring a variety of different things. Most commonly, trust refers to an
individual’s trust in other entities across levels of abstractions. As depicted in
Figure 2.3, trust can be discussed in the context of trust in another person, trust in
several other people within a group or organizations, trust in a specific company
or platform, or trust in the society in general.3 Below I review specific work on
3One can of course theoretically discuss trust originating from not the individual, but other
23
Figure 2.3: Multilevel perspective of trust.
trust at each level of abstraction.
Trust in another person (dyadic). Trust in another person, or commonly
referred to as dyadic trust, is one of the most common conceptualizations of
interpersonal trust [Larzelere and Huston, 1980, Berg et al., 1995]. Dyadic trust
refers to the decision to trust or the assessment of trustworthiness of a specific
individual, as opposed to people in aggregate [Larzelere and Huston, 1980].
From a social exchange theory point of view, dyadic trust is trust involved
in direct exchange. One key advantage of studying dyadic trust is that it is
the “minimal” meaningful unit of interaction that involves trust, providing
conceptual simplicity to investigate factors that affect trust.
The field of social psychology and economics both discuss dyadic trust
heavily, but using different methods. In psychology, dyadic trust is most com-
monly measured with techniques in psychometrics, by developing interper-
entities, such as how a company trusts another company. However, such discussion is less
common in prior literature and also not the focus of this dissertation.
24
sonal trust scales. For example, the Dyadic Trust Scale measures dyadic trust
with agree/disagree items including “My partner is primarily interested in
his (her) own welfare”, and “My partner is perfectly honest and truthful with
me” [Larzelere and Huston, 1980]. Such measures of dyadic trust has shown
that dyadic trust correlates with love and intimacy of self-disclosure [Larzelere
and Huston, 1980]. The work presented in Chapter 4 in this dissertation follows
this tradition and investigates dyadic trust on Airbnb using a interpersonal trust
scale adapted to the domain of lodging.
In economics, dyadic trust is most commonly measured through trust
games [Berg et al., 1995]. Initially developed by Berg et al., the trust game
is an economic decision-making game that has built-in risk dynamics to allow
for the measurement of trust and trustworthiness [Berg et al., 1995] (I expand on
details of the trust game in Section 2.2.2). Since then, the trust game has become
a useful instrument to study trust in dyadic settings in and outside of the field of
economics, and often in experiments [Kiyonari et al., 2006,Kuwabara, 2015,Abra-
hao et al., 2017, Yamagishi et al., 2009, Resnick and Zeckhauser, 2002, Qiu et al.,
2018, Merrill and Cheshire, 2017]. For example, Kuwabara found that reputation
systems can reinforce trust and trustworthiness through experiments using trust
games [Kuwabara, 2015]. In another experiment, it was found that reputation can
overcome homophily in establishing trust, again using trust games to measure
dyadic trust [Abrahao et al., 2017]. Finally, in an investigation on how bio sig-
nals affect trust, trust games were used to show that elevated heartrates reduce
trust [Merrill and Cheshire, 2017].
Trust within groups/organizations. While it is useful to study trust in direct
exchange (dyadic trust), trust in generalized exchange is also important to under-
25
stand. One common form of generalized social exchange is through social groups
or organizations. Trust within groups or organizations4 refers to trust in people
in aggregate in the group or organization, as opposed to a specific individual.
In particular, my work in Chapter 5 focuses on predicting trust within groups
quantitatively.
Work in social psychology, communication, and management science has
examined trust within groups or organizations. In social psychology, trust
within groups was found to stem from basic properties of the group such as its
size [Brewer, 1991]. For instance, experiments have shown that people identify
more strongly with smaller groups [Simon and Brown, 1987]. Trust within groups
was also found to indirectly influence group task performance [Dirks and Ferrin,
2002]. In communication, trust within groups was found to correlate with rule
following behaviors [Walther and Bunz, 2005]. However, trust within groups can
also be fragile and temporal when the group is formed around a common task
with a finite life span [Jarvenpaa and Leidner, 1999, Meyerson et al., 1996].
In management science, an important theoretical framework on organiza-
tional trust was proposed by Mayer et al. in 1995, which took an integrative
approach by synthesizing prior work on trust from multiple disciplines [Mayer
et al., 1995, Schoorman et al., 2007]. Three dimensions of trust, ability, benevo-
lence, and integrity, were proposed. Ability refers the competency or technical
skills of the trustee; benevolence refers to the extent the trustee is believed to
want to do good to the trustor; and integrity refers to the belief that the trustee
will adhere to a set of principles that the trustor finds acceptable [Mayer et al.,
4I use the term trust within groups or organizations here to distinguish better from trust
in companies/platforms/government below, where the trustor is not part of the other entity.
However, in Chapter 5, the term trust in groups is used to refer to trust within groups for
simplicity.
26
1995]. Though originally developed in the organizational context, in my work
I adapt these dimensions in a variety of settings to measure trust in surveys,
including trust in platforms in Chapter 3, dyadic trust in Chapter 4, and trust
within groups in Chapter 5. Using such framework, trust in management within
organizations has been shown to positively relate to employee’s ability to focus
attention on value-producing activities [Mayer and Gavin, 2005]. In addition,
trust within teams has been found to be positively related to perceived task
performance, team satisfaction, relationship commitment, and negatively related
to stress and tardiness [Costa et al., 2001, Bijlsma and Koopman, 2003,Colquitt
et al., 2007].
Finally, as people more commonly engage with social groups online, new
opportunities arise to study trust within groups online, especially how group net-
work structure relates to trust. Several studies in Human-Computer Interaction
began to examine trust in social groups in social networks. In one qualitative re-
search on Facebook buy-and-sell groups, trust was found to be fostered through
mechanisms such as closed membership and sanctioning [Moser et al., 2017]. In
addition, network density was also found to lead to higher trust in buy-and-sell
groups [Holtz et al., 2017]. My work builds on existing work to understand and
predict trust within Facebook groups in Chapter 5.
Trust in companies/platforms/government. Another important level of trust
is trust in companies, platforms, and the government. In contrast to trust within
groups or organizations where the trustor is part of the group, the trustor here is
not a part of the organization.
Most commonly, marketing literature discusses trust in companies and
brands, while political science literature discusses trust in the government. In
27
marketing, trust in brand was found to determine consumer satisfaction and
loyalty, both in terms of purchase behavior and attitudes [Chaudhuri and Hol-
brook, 2001, Delgado-Ballester and Luis Munuera-Alemán, 2001]. In addition,
consumer trust in online commerce platforms has also received research at-
tention, especially in the early 2000s when they first emerged [Corritore et al.,
2001, Gefen and Straub, 2004]. In my work, Chapter 3 presents a study of trust in
online commerce platforms. In political science, trust in government, or public
trust, is a crucial research question especially important for the study of democ-
racy [Hardin, 1999, Chanley et al., 2000]. Prominent political scientists such as
Francis Fukuyama argue that public trust fosters social and economic prosper-
ity [Fukuyama, 1995]. However, in past few decades, literature has tracked a
steady decline of generalized trust in society [Pew, 2019, Miller, 1974, Chanley
et al., 2000]. Relatedly, trust in media and news has also been on decline, espe-
cially as concerns over misinformation grow [Kohring and Matthes, 2007, Gun-
ther, 1988, Allcott et al., 2019, Flintham et al., 2018, Grinberg et al., 2019, Flintham
et al., 2018].
Finally, as technological platforms play an increasingly important role in
every aspect of people’s lives, trust, or rather, distrust in technological platforms
has gained significant attention, especially in the areas of Human-Computer
Interaction, Computer Supportive Collaborative Work, Science and Technology
Studies, and law and policy research. In particular, distrust in social networking
sites such as Facebook has brewed over the years, exacerbated by the emotional
contagion study [Kramer et al., 2014], and Cambridge Analytica scandal5 [Dwyer
et al., 2007, Fogel and Nehmad, 2009, Lankton and McKnight, 2011]. Opacity
and bias in algorithms that are ubiquitous on digital platforms are likely to
5https://www.nytimes.com/2018/03/17/us/politics/cambridge-analytica-trump-
campaign.html
28
further erode trust in platforms [Barocas and Selbst, 2016, Buolamwini and
Gebru, 2018, Datta et al., 2018, Bolukbasi et al., 2016, Eckhouse et al., 2019].
Taken together, the decline of public trust, trust in media and news organiza-
tions, and distrust in technological platforms, are among the major challenges
that the world faces around trust today. It is important for future research on
trust to address these challenges.
Generalized trust and individual propensity to trust. Trust in the broadest
sense measures trust in other people in the society in general and in aggregate.
Trust in this context is commonly referred to as “generalized trust” [Nannestad,
2008], a “propensity to trust” [Ferguson and Peterson, 2015], or a “disposition to
trust” [Wu et al., 2010].6 Generalized trust is considered as a stable individual
trait, similar to that of personality [Rotter, 1967]. Previous work has also studied
cross-country differences in generalized trust, where social polarization in the
form of income inequality and ethnic diversity was found to reduce trust [Bjørn-
skov, 2007]. Therefore, it is useful as a control variable in the study of trust
in other contexts. For example, in Chapter 5, generalized trust was used as a
control variable in the prediction of trust within groups and was found to have
significant predictive value.
Exact definitions of generalized trust and individual propensity to trust are
slightly different, although they are largely considered to be very similar. Gener-
alized trust is usually measured with the trust question in General Social Survey
(GSS) or the World Values Survey (WVS), “Generally speaking, would you say
that most people can be trusted or that you can’t be too careful in dealing with
people?” Such measure has been subjected to criticism that that the interpretation
6Technically, trust in this context refers to the average level of perceived trustworthiness of
others in the society.
29
of “most people” could vary widely person to person, therefore can be unreli-
able with stochastic errors [Nannestad, 2008]. In addition, generalized trust has
been shown to have little predictive power when it comes to specific individual
behaviors in games that involve trust [Glaeser et al., 2000]. On the other hand,
individual propensity to trust has followed a psychometric tradition, measured
with instruments such as the Rotter Interpersonal Trust Scale (ITS) [Rotter, 1967].
The 25 agree/disagree items in ITS ask about perceived trustworthiness across a
variety of different roles in society, and take the average of the responses as the
trust composite score. These agree/disagree items include “Parents can usually
be relied upon to keep their promises.”, “Most elected public officials are really
sincere in their campaign promises”, etc.
One line of work suggests that individual propensity to trust is rooted in life
experiences and societal norms [Rotter, 1971,Bowlby, 1969,Ainsworth et al., 2015].
Propensity to trust is considered as an important variable in the development
of healthy family relationships and personalities in children [Rotter, 1967] It has
also been associated with being older, married, having higher family income and
college education and living in a rural area, but not with gender [Taylor et al.,
2007, Paxton, 2007]. A propensity to trust is also related to other personal traits,
such as risk-taking [Cook et al., 2005], feelings of social support, and group
loyalty [Barrera Jr and Ainlay, 1983, Van Vugt and Hart, 2004].
A multilevel perspective. As reviewed above, when we talk about trust,
it can refer to trust in entities across different levels of abstractions. Different
levels of abstractions, as well as different traditions and language of disciplines
created challenges for trust research. To get a holistic view of trust, it is important
to synthesize various viewpoints and examine how trust at different levels of
30
Figure 2.4: The integrative organizational trust model proposed by Mayer
et al. 1995
abstractions interact with each other. For example, how does trust in platforms
affects trust in information shared by another person?
Mayer et al. 1995 proposed an integrative approach to study trust, which has
been highly impactful [Mayer et al., 1995]. In particular, as shown in Figure 2.4,
Mayer et al. emphasized how individual propensity to trust can mediate the
relationship between impressions of perceived trustworthiness and the decision
to trust.
At the same time, with increasing digital mediation in our social exchange,
the integrative model proposed by Mayer et al. needs further development.
Interpersonal exchange takes place embedded in social networks (e.g., Facebook,
LinkedIn, etc.). In addition, increasing algorithmic mediation in online platforms
is also not accounted for in Mayer et al.’s model. A new framework is needed
to address the challenges in trust brought by social networks and algorithmic
31
mediation. To do so, in Section 2.3, I develop a framework of “networked trust”
to incorporate these new factors in digitalized exchange.
2.2.2 Methodological Note on Measuring Trust
When developing new research projects on trust, deciding how to measure trust
is often challenging. Here I offer a brief methodological note on the measurement
of trust to facilitate future research. There are mainly three ways that prior
literature used to measure trust: (1) surveys; (2) real-world behavior proxies; and
(3) lab-based trust games. Each method has strengths and weaknesses, which we
discuss briefly below.
Surveys are among the most widely used instrument to measure trust. We can
use surveys to ask the willingness for the trustor to make a trust decision, though
usually in a hypothetical setup. For example, Chapter 4 used a survey-based
experiment to ask about the decision to stay with a particular host on Airbnb. In
another example, Kizilcec used a questionnaire to study how transparency of an
algorithm affects trust [Kizilcec, 2016]. However, such measures of trust have
limited construct validity, as self-report answers in hypothetical decision-making
scenarios can be unreliable.
On the other hand, surveys can also measure perceived trustworthiness,
which is a subjective assessment and an internal state of the trustor. The World
Values Survey has a few survey questions tracking generalized trust (perceived
trustworthiness of other people in society) since the 1970s across different coun-
tries. One of the most well-known survey question on trust is: “Generally
speaking, would you say that most people can be trusted or that you need to be
32
very careful in dealing with people?” [WVS, 2018]. Respondents can choose one
of the options between “Most people can be trusted”, “Need to be very careful”,
and “Don´t know”. This measure and its variations have also been widely used
in political science. For example, surveys on political trust ask “How much of
the time do you think you can trust the government in Washington to do what is
right?” [Miller, 1974].
Finally, reputation systems can be broadly considered as a survey-based
method to establish a proxy measure for trustworthy behaviors. After an ex-
change has taken place, a survey is issued to exchange partners, asking them
to provide feedback and rating towards each other. Then the reputation system
aggregates the feedback and ratings and displays them for future interactions.
It is important to note that reputation systems have inherent limitations such
as bias (racial bias, selection bias, and positivity bias), and are often subject to
problems including fake reviews and cold start [Fradkin et al., 2015, Resnick and
Zeckhauser, 2002, Zervas et al., 2015, Ott et al., 2011]. A more detailed review of
prior work on reputation systems can be found in Section 2.3.
Real-world behavior proxies have higher construct validity in measuring
trust compared to survey-based method. For example, when someone decides to
fund a crowdsourcing campaign [Mitra and Gilbert, 2014], we can consider the
decision or behavior to fund as trust in-situ. Such behavior proxies are becoming
increasingly available on digital platforms, such as the decision to meet someone
offline through online dating platforms, the decision to book an Airbnb, the
decision to purchase something online, etc. These behavior proxy measures are
being used in trust research to show the impact of trust on real-world outcomes
at scale. Abrahao et al. used actual Airbnb behavior data to validate findings
33
about reputation and trust [Abrahao et al., 2017]. In this dissertation, real-world
behavior proxies have also been an important part of research findings — the
work presented in Chapter 3 leveraged eBay sales data, and the work presented
in Chapter 5 leveraged Facebook groups behavioral data [Ma et al., 2019a] such
as group densification (a behavior proxy for the decision to trust).
At the same time, behavior proxies can be valuable for addressing problems
of bias in online marketplaces and reputation systems. When people make
decisions about trust, they rely on information that is available to them to assess
the perceived trustworthiness of another party. Bias exists when people make
up such impressions (more details on bias are available in Section 2.4). As
digital platforms accumulate data about the actual behavior of the trustee (e.g.,
whether a loan is repaid), behavior proxy data can be used to assess the actual
trustworthiness of the trustee, reflecting ground truth rather than impressions.
According to signaling theory [Donath, 2007], behavior proxies can be considered
as “assessment signals” (something inherent to the signal itself connects it to
the quality it indicates) as opposed to “conventional signals” (something not
inherently reliable, but are kept so through conventions). There are opportunities
to make behavior proxy data available to help address the bias problem in online
marketplaces. For example, tracking and displaying the real response time of an
Airbnb host, or a business on Facebook messenger can establish trust in the true
level of responsiveness, rather than claims.
However, assessing the trustworthiness through behavior proxy also faces
challenges. By definition, trustworthiness can only be measured after an inter-
action has taken place, while the decision to trust has to be made before the
interaction. Sometimes, long periods of time pass before we learn the true behav-
34
ior of the trustee. For example, in the case of funding startups, it may take at least
a few years before the investors learn about the true level of trustworthiness of
the team being funded. Second, there is no guarantee that future trustworthiness
will be consistent with prior behavior (e.g., Ponzi scheme). Repeated measures
can be one way to gain more confidence that the trustee will behave consistently
to our assessment. But due to limitations in time, repeated measures can be chal-
lenging to obtain. Finally, behavior proxy can be hard to collect or interpret when
lacking contextual knowledge. For example, in the case of Airbnb, it is very hard
to collect data about the actual condition of the listing, or about whether there
are hazards such as hidden cameras inside the listings. In the case of Uber, when
the GPS data shows that the driver took a wrong route, is it because the driver is
not responsible, due to unpredictable conditions on the road and the driver is
reacting responsibly? As such, finding better behavior proxies to measure trust
remains future work.
Lab-based trust games are another important instrument to measure trust.
Although the work presented in this dissertation relies mostly on surveys and
behavior proxies to measure trust, trust games can be useful in studying how
certain factors affect trust. Trust games are economic games designed to abstract
risks and rewards in real-world exchanges to allow for more controlled study
of trust, commonly in experimental setups. By abstracting contextual risks into
numerical payoffs, all measures of trust can be reduced to a numerical based
decision — making it easy to compare across conditions.
Specifically, the original trust game was developed by Berg in 1995 — an
investment game that uses the amount of investment to denote the level of trust
in the participant’s partner [Berg et al., 1995]. The setup of the trust game is
35
collaboration through risk taking: One participant is given a certain amount of
money and can choose an amount between zero and the total amount to give to
the partner in the experiment (measures trust). The experimenter will triple the
amount that is sent to the other participant, and the other participant can decide
how much between zero and the total endowed amount to send back to the first
participant (measures trustworthiness). The more one trusts the partner, the
more amount in the beginning one should sent, as both parties together achieve
maximum reward that way. However, the risk of the second participant not
returning any amount makes it a prerequisite for the first participant to have
trust. Results show that trust indeed exists. In one of the conditions, the first
participant sent on average $5.36 out of $10, and the other participant returned
on average $6.46 out of $15 endowed.
Many subsequent research adopted trust games. For example, Kiyonari et
al. found that trust does not beget trustworthiness in one-shot games [Kiyonari
et al., 2006]. Abrahao et al. used trust games to study the relationship between
homophily, reputation and trust [Abrahao et al., 2017]. Finally, Glaeser et al.
measured trust and trustworthiness using both the survey-based method, as
well as trust games, and found that standard attitudinal survey questions about
trust predicts trustworthy behavior in trust games much better than they predict
trusting behavior [Glaeser et al., 2000]. However, there are also key limitations to
using trust games to measure trust. For example, according to a meta-analysis of
162 trust games across 35 countries [Johnson and Mislin, 2011], behavior in trust
games is sensitive to the specific setup of the game. Factors such as the multiplier
set by the experimenter, whether the play is with a simulated counterpart, or
whether subjects play both roles in the experiment all affect trust.
36
2.3 Online Trust and Networked Trust
A big focus of this dissertation is to address new questions about trust that arise
as a result of the digitalization of exchange. In this section, I first review prior
work on online trust, primarily focusing profiles and reputation systems. Then, I
briefly trace different waves of the digitalization of exchange and how they affect
social structure.
Finally, I propose the framework of “networked trust” to reason about trust in
the increasingly connected age, with three focuses: cues, networks, and algorithms.
2.3.1 Online Trust
Online trust is not an entirely new phenomenon — work since early 2000s
primarily focused on how two mechanisms produce online trust: (1) profiles; (2)
reputation systems.
Profiles are an important part of many online systems [Uski and Lampinen,
2014]. Due to the computer-mediated nature of online systems, profiles provide
an identity for users that persists over time and the myriad of interactions on
the site [boyd and Ellison, 2007]. In online dating sites, profiles provide self-
disclosure that attracts interests of other users, while limiting the risks associated
with outright deception [Gibbs et al., 2010, Toma et al., 2008].
One way to understand how profiles establish trustworthiness is the Profile
as Promise framework [Ellison and Hancock, 2013]. Profile as Promise uses the
notion of a promise to characterize the function of a profile. In this view, the
37
profile is a psychological contract between the profile holder, and the viewer that
future interactions (e.g. with a date, a car driver, or an Airbnb host) will take place
with someone who does not differ fundamentally from the person represented
in the profile. The notion of a promise has been successfully applied to various
contexts that require good faith to operate, including online dating [Ellison and
Hancock, 2013] and job hunting [Rousseau and Greller, 1994].
Reputation systems represent another well studied mechanism for online
trust. A cornucopia of work on reputation systems, focused on reputation in
online commerce platforms, showed the usefulness as well as the limitations of
this mechanism in establishing online trust [Jøsang et al., 2007].
At its core, reputation systems provide aggregated information about the
trustworthiness of potential exchange partners based on how they have behaved
in the past. However, selection bias and positivity bias in reputation systems
limit the accuracy and usefulness of reputation systems.
Past research on reputation systems has focused on how reputation systems
change people’s behavior and their limitations.
In terms of how reputation systems change people’s behavior, Resnick et
al. are among the firsts to study trust and reputation systems [Resnick et al.,
2000, Resnick and Zeckhauser, 2002, Resnick et al., 2006]. Through a controlled
field experiment, Resnick et al. found sellers with a strong reputation receive a
price premium reflecting higher trust on eBay [Resnick et al., 2006]. Kuwabara
found that reputation systems can reinforce generalized trust and trustworthiness
— recalling one’s reputation on eBay made participants behave more trustworthily
in the relevant roles [Kuwabara, 2015]. Yamagishi et al. conducted online trading
38
experiments to see how reputation systems can solve the “lemons problem” for
online marketplaces by encouraging more trustworthy behavior [Cook et al.,
2009]. A most recent experiment using Airbnb users as participants showed that
reputation systems can increase trust between dissimilar users, hence mitigating
bias created by homophily [Abrahao et al., 2017].
In terms of the limitations of reputation systems, through empirical analysis
on eBay, Resnick et al. found that while only half of the buyers provide feedback
(selection bias), the overwhelming majority of feedback is positive (positivity
bias) [Resnick and Zeckhauser, 2002]. Similar selection bias and positivity bias
were observed on Airbnb [Fradkin et al., 2015,Zervas et al., 2015] and freelancing
marketplace oDesk [Filippas et al., 2018]. Non-reviewers on Airbnb were found
to have worse experiences than reviewers [Fradkin et al., 2015]. As a result,
“every stay is above average” on Airbnb [Zervas et al., 2015] — nearly 95%
of Airbnb properties boast an average user-generated rating of either 4.5 or 5
stars (the maximum).
In addition to selection bias and positivity bias, reputation systems are also
vulnerable to fake reviews. There is an arms race between computational fake
review detection and generation. In 2011, simple linguistic features such as
n-grams and LIWC features [Pennebaker et al., 2001] were sufficient to detect
fake reviews generated by humans [Ott et al., 2011]. In 2017, with advancements
in deep learning, AI can generate fake Yelp reviews that are “not only evade
human detection, but also score high on ‘usefulness’ metrics by users” [Yao et al.,
2017b]. Finally, reputation systems suffer from the cold start problem, which
refers to the fact that there are no reviews when systems initialize [Tavakolifard
and Almeroth, 2012].
39
2.3.2 Networked Trust
Although online trust is not a totally new phenomenon, new challenges continue
to arise as digitalized exchange platforms evolve. There is a narrative that there
has been an decline in trust in the past few decades, especially trust in the
government [Pew, 2019, Miller, 1974, Chanley et al., 2000]. According to a 2019
survey by Pew Research Center, public trust in the government remains near
historic lows. “Only 17% of Americans today say they can trust the government
in Washington to do what is right ‘just about always’ (3%) or ‘most of the
time’ (14%).” [Pew, 2019]. However, at the same time, people engage in social
exchange that require high levels of trust at scale — interactions that were totally
unimaginable just a few years back (e.g., staying at strangers’ homes in the
case of Airbnb). At the same time, even though traditional community-based
organizations have seen decline in memberships (e.g., bowling leagues [Putnam,
1995]), online communities have exploded [Kraut and Resnick, 2012, Adamic
et al., 2008, Zhang et al., 2007, Lampe et al., 2006]. Therefore, it is not informative
to simply discuss whether there has been a “decrease” or “increase” in trust in
the society, but rather, how trust structurally shifts from centralized institutions
towards trust in networked environments.
In my view, fundamental changes to trust, driven by the digitalization of
exchange, took place in three waves. (1) the digitization of the exchange of
goods (e.g., Amazon, eBay); (2) the digitization of social relationships (e.g., Face-
book, Twitter, LinkedIn, Instagram); and (3) the digitalization of the exchange of
resources through sharing economy platforms (e.g., Airbnb, Uber, Lyft).
In the first wave, people have to trust each other primarily with their financial
resources. Mechanisms to reduce risks, such as third-party payment systems and
40
reputation systems have overall successfully addressed the trust problem. New
questions are still being raised, though, as online commerce platforms move
towards more mobile-based where user-generated content play a more important
role in establishing trust.
In the second wave, people have to trust others with their personal information.
The social nature of these online platforms, primarily social network sites (SNSs),
means that profiles play a central role in establishing trust. As people rely on on-
line profiles rather than signals that are available in face-to-face communications,
there are a lot of questions about how cues in Computer-Mediated Communica-
tion (CMC) establish trust. At the same time, through SNSs, people are able to
maintain more weak ties, leading to “networked individualism”. We shift away
from small, isolated, and tightly-knit groups to networked individuals [Rainie
and Wellman, 2012, Granovetter, 1973]. At the same time, because SNSs make
social relationships visible and data available, network analysis can reveal how
network structures regulate information cascade [Cheng et al., 2014] and predict
social and economic outcomes [Easley et al., 2010, Backstrom et al., 2006].
In the third wave, people have to trust others not only with their financial
resources, personal information, but also physical safety. The wide adoption of
sharing economy platforms is a relatively recent phenomenon, and they blur the
line of social and economic exchange [Hamari et al., 2016, Ikkala and Lampinen,
2015, Lampinen and Cheshire, 2016]. Both profiles and reputation systems play
an important role in establishing trust in sharing economy. Compared to SNSs,
sharing economy platforms may involve higher risks (physical safety). At the
same time, people tap into the benefits of conducting exchange with strangers
on sharing economy platforms, rather than with their social ties, weak or strong.
41
Therefore, self-disclosure may be even more important to establish initial trust
that is required for interactions on sharing economy platforms.
In addition to these three waves of the digitalization of exchange, another
factor is playing an increasingly important role in mediating social exchanges
— algorithms. Online platforms deploy algorithms through personalization
and recommendation systems to optimize for specific engagement goals, and
these algorithms affect how people interact and trust each other. For example,
algorithms can match people in online dating sites [Hutson et al., 2018], curate
what people see and impact how they make sense of social relationships on
Facebook news feed [Eslami et al., 2015], and select to hide or display specific Yelp
reviews [Eslami et al., 2019]. Algorithmic mediation creates more uncertainty in
social exchange, potentially hindering interpersonal trust.
In short, the three waves of digitalization of exchange and the increasing
algorithmic mediation create new questions about trust and call for a new way
of thinking and reasoning about trust going forward.
Networked trust is such framework for thinking about trust online going
forward. Borrowing from the term of “networked individualism” that describes
the drastic shift in social structure brought by digital platforms [Rainie and
Wellman, 2012], “networked trust” describes interpersonal trust in digitalized
social exchange where cues, networks, and algorithms play important roles.
Networked trust has three focuses: (1) cues in Computer-Mediated Communi-
cation; (2) embeddedness in social networks; and (3) increasing mediation by
algorithms.
First, cues in Computer-Mediated Communication refers to interpersonal
42
trust where the initial social exchange takes place online and not face-to-face,
following the rules of Computer-Mediated Communication. This aspect of
networked trust is relatively well established, as I reviewed before in Section 2.3.1
on online trust. Two chapters of this dissertation, Chapter 3 and Chapter 4 expand
on the focus on cues in networked trust — analyzing image and language cues
at scale in commerce and sharing economy platforms respectively.
The second focus of networked trust, embeddedness in social networks,
refers to the fact that our social exchange is embedded in social networks. For
example, on LinkedIn, when users view another person’s profile, LinkedIn can
reveal to users information of the network, such as how two people are connected
or who their common connections are. Chapter 5 expands on the networks focus
of networked trust, using the context of Facebook groups.
Finally, the third focus of networked trust, increasing mediation by algo-
rithms, refers to the increasing control by algorithms on how and with whom
we conduct exchange with. This focus is still a nascent area of research, closely
related to the ongoing research on the issues of algorithmic fairness, accountabil-
ity, and transparency (FAT*). In Chapter 6, I discuss future work that focuses
on understanding how algorithms mediate exchange and impact interpersonal
trust.
2.4 Bias and Distrust
In the last section of the chapter, I review literature on the flip side of trust —
bias and distrust. As the process of forming trust is usually subjective, bias is
closely related to trust. This section reviews work demonstrating four different
43
types of biases, including racial, gender, status, and algorithmic bias. Given the
vast amount of literature on bias, I do not attempt to provide a comprehensive
review but rather focus on studies on marketplaces or closely related to trust.
Racial bias. Research has observed racial bias in labor markets through a
field experiment manipulating names on resumes [Lavergne and Mullainathan,
2004]. In the study, for resumes with similar credentials, white sounding names
received 50 percent more callbacks for interviews than African-American sound-
ing names across occupation, industry, and employer size. Racial bias is also
observed on Airbnb through a field experiment by manipulating names in online
profiles [Edelman et al., 2017]. Booking requests from guests with distinctively
African-American names are roughly 16% less likely to be accepted than identical
guests with distinctively white names.
In addition to names, visual information can also lead to racial bias online.
One experiment showed that photos in online marketplaces with a dark-skinned
hand holding the product lead to fewer and lower offers than photos with a
light-skinned hand, indicating lower trust [Doleac and Stein, 2013]. Finally, in
freelance marketplaces such as TaskRabbit, black workers receive significantly
fewer reviews and worse ratings [Hannak et al., 2017].
Gender bias. One of the most prominent gender bias in marketplaces is the
gender pay gap in labor markets. Women roughly earn 80 cents on average for
each dollar earned by men in the labor market [Blau and Kahn, 2007,Dubey et al.,
2017,Chamberlain, 2016]. Women sellers on eBay also receive less bids and lower
final prices in auctions compared to men counterparts [Kricheli-Katz and Regev,
2016].
44
Representation and participation bias are also present in gender. In 2012,
Vasilescu et al. provided empirical evidence showing that women on StackOver-
flow represent the minority of contributors and they participated less and earn
less reputation [Vasilescu et al., 2012]. Similar gender representation gap was
reported in editors of Wikipedia — a survey in 2010 found that fewer than 13%
of Wikipedia contributors were women [Antin et al., 2011].
Status bias. Bias with regard to socio-economic status has been observed
on sharing economy platforms. Research found that sharing economy is sig-
nificantly more effective in dense, high socio-economic status areas than in
low-socio-economic status areas [Thebault-Spieker et al., 2017]. For example,
lower socio-economic status areas have higher wait times on Uber; and workers
on TaskRabbit are less willing to travel to low socio-economic status areas for
work and charge higher prices in those neighborhoods [Thebault-Spieker et al.,
2017].
Status bias can also be context-induced. Contextual status can be induced by
power imbalance in specific situations. For example, in the context of Couchsurf-
ing, hosts have higher status compared to guests because the hosts provide free
lodging for the guests. As a result, contextual status differences arise, and hosts
receive higher ratings than guests in the same interactions, potentially due to the
mechanism of status-giving in power dependence theory [State et al., 2016].
Algorithmic bias. The last type of bias has only recently begun to be better
understood — algorithmic bias. In the 2016 paper, Big Data’s Disparate Impact,
Barocas and Selbst laid out how data mining can exhibit discrimination that
can be unintentional and hard to understand. As a result, current law and tech-
niques are inadequate in identifying and regulating the problem of algorithmic
45
bias [Barocas and Selbst, 2016].
Recent literature uncovers subtle racial and gender biases exhibited by algo-
rithms [Buolamwini and Gebru, 2018,Datta et al., 2018,Bolukbasi et al., 2016,Eck-
house et al., 2019]. For example, an audit of two standard benchmark facial
recognition datasets and three commercial systems revealed that the datasets
are overwhelmingly composed of lighter-skinned subjects and the commercial
systems have much higher rate of error for darker-skinned females than lighter-
skinned men (34.7% v.s. 0.8%) [Buolamwini and Gebru, 2018]. Bolukbasi et al.
also showed that word embeddings, a popular technique in natural language
processing, could exhibit gender stereotypes such as associating “receptionist”
to “female” [Bolukbasi et al., 2016].
Algorithmic bias raises the question: how might algorithmic bias affect trust?
Work in transparency and interpretability of machine learning examines bias and
trust in algorithms [Miller, 2018, Kizilcec, 2016, Eslami et al., 2015, Eslami et al.,
2019]. Another line of research is policy-driven [Levy and Barocas, 2017, Hutson
et al., 2018], outlining design and policy choices for platforms to make themselves
less conducive to discrimination by users to gain user trust. At the same time,
algorithmic bias might hinder interpersonal trust when the algorithms are crucial
in mediating social exchange relationships.
2.5 Summary
Gambetta once wrote:
“The importance of trust pervades the most diverse situations where
46
cooperation is at one and the same time a vital and a fragile com-
modity: from marriage to economic development, from buying a
second-hand car to international affairs, from the minutiae of social
life to the continuation of life on earth.” [Gambetta, 1988]
In this chapter, I reviewed prior work across many disciplines on trust in
“diverse situations”. By anchoring the definition of trust in social exchange theory,
the work presented in the rest of the dissertation focuses on addressing new
questions that the digitalization of exchange brings. Literature on different levels
of trust was reviewed, including trust at the individual level, trust in groups,
trust in platforms, trust in organizations, and trust at societal level. Finally, the
framework of “networked trust” was proposed, which has three focuses: (1) cues
in Computer-Mediated Communication; (2) embeddedness in social networks;
and (3) increasing mediation by algorithms. As trust is a subjective process, we
need to caution against biases, especially the ways that racial, gender, status,
and socio-economic biases can manifest subtly in algorithms that mediate social
exchange relationships.
47
CHAPTER 3
IMAGES OF TRUST: UNDERSTANDING IMAGE QUALITY AND TRUST
IN PEER-TO-PEER MARKETPLACES
3.1 Introduction
In this chapter and the chapter that follows, I present two case studies that
focus on the first focus of networked trust: trust in interactions where cues in
Computer-Mediated Communication play an important role.
Specifically, this chapter explores how image cues establish trust in the con-
text of second-hand peer-to-peer online marketplaces. As mentioned in Chap-
ter 2, online marketplaces represent the first wave of the digitalization of social
exchange, the digitization of the exchange of goods. Trust in the context of online
marketplaces was the earliest focus of the discussion on online trust. Reputation
systems were pioneered first in online marketplaces to facilitate online trust and
research has discussed their benefits and limitations [Resnick et al., 2000, Resnick
and Zeckhauser, 2002, Resnick et al., 2006, Cook et al., 2009].
However, as online commerce platforms continue to evolve and new tech-
nologies emerge, new challenges and opportunities about understanding trust
on online commerce platforms arise. Increasingly, these platforms are becoming
mobile-based, where user-generated images play an ever increasingly important
role, compared to stock imagery. Although earlier work has studied the role of
images in online marketplaces computationally [Goswami et al., 2011, Chung
et al., 2012], user-generated images are less well understood. At the same time,
advancements in computer vision (e.g., deep learning) present new opportunities
48
to understand image cues with more advanced techniques.
This chapter leverages such opportunities brought by advancements in com-
puter vision to present a study on how user-generated images establish trust in
second-hand peer-to-peer marketplaces.1 Trust here is discussed in the context
of trust in platforms, and was measured with a survey. The survey asked about
the perceived trustworthiness in hypothetical marketplaces where images vary.
At its core, images are representations of the products to be exchanged online.
Since it is impractical for a customer to inspect an item in person before purchas-
ing it, images are very useful to establish trust by setting expectations, reducing
uncertainty, and limiting information asymmetry in online marketplaces [Ak-
erlof, 1978]. “Good” product images achieve the goals of expectation setting and
uncertainty reduction effectively, and are likely to be successful in facilitating
exchange.
Traditionally, stock images are considered as the gold standard for “good”
images in online marketplaces, which are usually shot in professional studios
with high-end equipment. They present the products in the best possible light,
so to speak.
At the same time, as mobile-based peer-to-peer marketplaces such as
LetGo.com and Facebook Marketplace gain popularity, the definition of “good”
images may have evolved. These mobile-based applications enable easy listing
process: users can quickly snap a picture with their phone, type in a short de-
scription, and instantly post their offerings. The peer-to-peer nature of these
marketplaces also mean that sellers are often non-professionals, who either lack
1This work was published at WACV 2019 as Understanding Image Quality and Trust in Peer-to-
Peer Marketplaces [Ma et al., 2019b].
49
the expertise or motivation to provide a stock image style picture. As a result,
peer-to-peer marketplaces today often contain images of mixed-quality that are
user-generated rather than stock images.
In this context, it is important to understand how mixed-quality user-
generated images establish trust. Specifically, I ask the following questions:
Is it possible to have a definition of “good” user-generated product images that
human raters agree on reliability? In addition, is it possible to computationally
tell “good” and “bad” images apart accurately? And finally, how does image
quality affect market outcomes, especially trust and sales? The rest of the chapter
addresses these questions.
Specifically, this chapter makes the following contributions.
Annotating image quality. We curated a dataset of user-generated images
of two common second-hand products (≈75,000 images). Then an annotating
process was iteratively developed and we show that human raters can provide
reliable judgements on on image quality under this process. We further annotated
a third of the collected data with human-labeled quality judgment.
Modeling and predicting image quality. Previous work studied general
photo aesthetics, but we show existing models do not completely capture im-
age quality in the context of online marketplaces. In this work, we used both
black-box neural networks and interpretable regression techniques to model and
predict image quality (how appealing the product image appears to customers).
For the regression-based approach, we model the factors of the photographic
environment that influence quality with handcrafted features, which can guide
potential sellers to take better pictures. As a result, we developed a better
50
“I believe that products from 
these sellers will meet my 
expectations when delivered.” Tr
Agree 4 ust Gap!
3.75
3.5
3.25 Poor Quality High Quality Stock Images 
(Predicted) (Predicted)
Neutral 3
Figure 3.1: We study the interplay between image quality, marketplace
outcomes, and user trust in peer-to-peer online marketplaces.
Here, we show how image quality (as measured by a deep-
learned CNN model) correlates with user trust. User studies in
Sec. 3.6.2 show that high quality images selected by our model
out-performs stock-imagery in eliciting user trust.
understanding of how visual features impact image quality, while training a
convolutional network to predict image quality with decent accuracy (≈87%).
Marketplace outcomes: sales and perceived trustworthiness. Using our
learned quality model, we then showed that image quality scores are associated
with two different group outcomes: sales and perceived trustworthiness. Pre-
dicted image quality was associated with higher likelihood that an item is sold,
while high quality user-generated images selected by our model out-performs
51
Agreement
Figure 3.2: Samples of lowest-rated and highest-rated images from shoes
and handbags groundtruth.
stock imagery in eliciting perceived trust from users (see Figure 3.1). Our find-
ings can be valuable for designing better online marketplaces, or to help sellers
with less photographic experience take better product images.
3.2 Related Work
Image quality in online marketplaces. Previous work has shown that image-
based features such as brightness, contrast, and background lightness con-
tribute to the prediction of click-through rate (CTR) in the context of product
search [Goswami et al., 2011, Chung et al., 2012]. This line of work used hand-
crafted image features, but did not actually assess the image quality as dependent
variable. In another work on eCommerce, image quality was modelled and pre-
dicted through linear regression, and shown to be significant predictors of buyer
interest [Di et al., 2014]. However, the dataset is not available, nor any details
on the modelling methodology or model performance. Our work fills this gap
by contributing a large annotated image quality dataset, along with improved
model performance.
Automatic assessment of aesthetics. Another line of closely related work is
52
automatically assessing the aesthetic quality of images. Early work on aesthetic
quality frequently used handcrafted features [Datta et al., 2006, Li et al., 2010,
Bhattacharya et al., 2010]. More recent work focused on AVA (A Large-Scale
Database for Aesthetic Visual Analysis), a large-scale dataset containing 250,000
images with aesthetic ratings [Murray et al., 2012] and adapting deep learning
features [Lu et al., 2014] to improve prediction accuracy. In addition, aesthetics
quality has been shown to predict the performance of images in the context
of photography, increasing the likelihood that a picture will receive “likes” or
“favorites” [Schwarz et al., 2016]. Although aesthetic quality is a fundamental
and important feature of imagery, images online marketplaces belong to another
visual domain compared to most of the images in this line of work. We show
that aesthetic quality models do not completely capture product image quality.
Web aesthetics and trust. A large body of work in human computer inter-
action has investigated the link between the marketplace website’s aesthetics
and perceived credibility. For example, previous work has shown low level
statistics such as balance and symmetry correlate with aesthetics [Zheng et al.,
2009, Michailidou et al., 2008]. More recently, computational models have been
developed to capture visual complexity and colorfulness of website screenshots
to predict aesthetic appeal [Reinecke et al., 2013]. However, since product images
take up majority of the space in online marketplaces, inadequate attention has
been paid to how the image quality, rather than interface quality, impact user
trust. Our work focuses on product images as the most salient visual element of
online marketplaces to study how they contribute to user trust.
Our work is concerned with user trust. Perceptions of user trust is important
for several reasons: trust influences loyalty, purchase intentions [Hong and Cho,
53
2011], retention [Sun, 2010], and is important for the platform’s initial adoption
and growth [Choi and Mai, 2018]. This is why we take a “social science” approach
when soliciting trust judgments in Sec. 3.6.2.
Domain adaption: matching street photos to stock photos. One might won-
der why marketplaces do not simply retrieve the stock photo of the product
being depicted and display it instead. There are some problems with this ap-
proach. First, in used goods markets, stock photos do not depict the actual item
being sold and are generally discouraged [Bland et al., 2007]. Second, stock
image retrieval is a computationally challenging task, with state-of-the-art meth-
ods [Wang et al., 2016] achieving around 40% top-20 retrieval accuracy on the
Street2Shop [Kiapour et al., 2015] dataset. Our work also contributes to the
dataset of “street” photos taken by different users using a variety of mobile
devices in varying lighting conditions and backgrounds.
3.3 Datasets
As shown in previous work [Di et al., 2014], image quality matters more for
product categories that are inherently more visual (e.g., clothing). Thus, in
our development of the dataset, we focus on the shoes and handbags categories.
These two categories are among the most popular goods found on secondhand
marketplaces and are visually distinctive enough to pose an interesting computer
vision challenge.
There are two sources for the data used in this work: LetGo.com, and eBay.
We focused on the publicly available LetGo.com images for creating the hand-
annotation dataset, and used private data from eBay to test the relationship
54
between image quality and marketplace outcome — sales.
3.3.1 LetGo.com
LetGo.com is a mobile-based peer-to-peer marketplace for used items, similar to
Craigslist. Potential buyers can browse through the “listings” made by sellers
and contact the seller to complete the transaction out of the platform.
We collected product images data for two product categories, shoes and hand-
bags. We crawled the front page of LetGo.com every ten to thirty minutes for a
month, filtering the listings by relevant keywords in the product listing caption.
For shoes, we used the keywords “shoe,” “sandal,” “boot,” or “sneaker” and
collected data between November to December 2017 (66,752 listings containing
133,783 images in total). For handbags, we used the keywords “purse” or “hand-
bag’ and collected data between April to May 2018 (29,839 listings containing
44,725 images in total).
3.3.2 eBay
To understand whether image quality impacts real world outcomes, we partnered
with eBay, one of the largest online marketplaces.
We collected data for listings on eBay in our two product categories, shoes and
handbags, including the product images, meta-data associated with the listing,
as well as whether the listing had at least one sales completed before becoming
expired. We sampled data based on the date on which the listing expired (during
55
May 2018). We also down-sampled the available listings to create a balanced set
of sold and unsold listings. In summary, this dataset included 66,000 sold and
66,000 unsold listings for shoes, and 16,000 sold and 16,000 unsold listings for
handbags.
To evaluate our model’s generalizability in a real-world setting, we train
models on publicly available LetGo.com data and test the relationship between
predicted image quality and sales on eBay.
3.4 Annotating Image Quality
We collected ground truth image quality labels for our LetGo.com dataset (LetGo
below) using a crowdsourcing approach. We designed the following task and
issued it on Amazon Mechanical Turk, paying $0.8 per task. Each task contained
50 LetGo images randomly batched, and at least 3 workers rated each image. For
each image, the worker was asked to rate the image quality on a scale between 1
(not appealing) to 5 (appealing). In total, we annotated 12,515 images from the
shoes category and 12,222 from the handbags category. Only LetGo data was used
for this task.
An important consideration is the difference between product quality and
photographic quality. In this survey, we are primarily interested in what the
merchant can do to make their listings more appealing, so it is important that
workers ignore perceived product differences. To help prime our workers along
this line of thinking, we added two text survey questions spaced throughout
each task with the following prompt: “Suppose your friend is using this photo to
sell their shoes. What advice would you give them to make it a better picture? How can
56
Standardized score distributions Handbags
5000 2500
0 2 0 2 0
5000 Shoes
5000 2500
0 2 0 2 0 Negative Neutral Positive
Scores
Figure 3.3: Left: standardized score distributions for filtered images on
shoes and handbags categories. Right: final ground truth labels,
showing a fairly even distribution.
the seller improve this photo?” This task also slows workers down and forces them
to carefully consider their choices.
After collecting the data, we first standardized each worker’s score distribu-
tion to zero mean and unit variance to account for task subjectivity and individual
rater preference. Then, we filtered out images where the standard deviation of
all rater scores was within the top 40%, retaining 60% of the original dataset.
This removed images where annotators strongly disagree. This filtering also im-
proved the inter-rater agreement as measured by an average pairwise Pearson’s
ρ correlation across each rater from 0.34 ± 0.0046 on the unfiltered shoes data to
0.70 ± 0.0031 on the filtered shoes data. This shows our labeling mechanism can
reliably filter images with low annotation agreement.
Finally, we discretized the scores into three image quality labels: good, neutral
and bad. To do this, we rounded the average score to the nearest integer, and
took the positive ones as good, zeros as neutral, and negative ones as bad. The
resulting shoes and handbags datasets are roughly balanced; see score distributions
in Figure 3.3 and example images in Figure 3.2.
57
Frequency Frequency
3.5 Modeling Image Quality
After collecting ground truth for our dataset, we study what factors of an image
can influence perceived quality. This turns out to be nontrivial. Shopping
behavior is complicated, and customers and crowd workers alike may have
intricate preferences, behaviors, and constraints.
We attempt to model image quality in two ways: first, we train a deep-learned
CNN to predict ground truth labels. This model is fairly accurate, allowing us
to approximate image quality on the rest of our dataset, but as a “black box”
model it is largely uninterpretable, meaning it does not reveal what high-level
image factors lead to high image quality. Second, to understand quality at an
interpretable level, we then use multiple ordered logistic regression to predict
the dense quality scores. This lets us draw conclusions about what photographic
aspects lead to perceived image quality.
Prediction. For our model, we use the pretrained Inception v3 network ar-
chitecture [Szegedy et al., 2016] provided by PyTorch [Paszke et al., 2017]. Our
task is 3-way classification: given a product image, the model predicts one
of {0,1,2} for bad, neutral, and good respectively. To do this, we remove the
last fully-connected layer and replace it with a linear map down to 3 output
dimensions.
Image quality measurements are subjective. Even though our crowdsourcing
worker normalization filters data where workers disagree, we want to allow
the model to learn some notion of uncertainty. To do this, we train using label
smoothing, described in § 7.5 of [Goodfellow et al., 2016]. We modify the negative
log likelihood loss as follows. Given a single input image, let xi for i ∈ {0,1,2} be
58
the three raw output scores (logits) from the model and let x̂i = logexpxi/∑ j expx j
be the output log-probabilities after softmax. To predict class c ∈ {0,1,2}, we use
a modified label smoothing loss, `(x,c) =−(1−ε)x̂c−ε ∑i6=c x̂i for some smoothing
parameter ε , usually set to 0.05 in our experiments. This modified loss function
avoids penalizing the network too much for incorrect classifications that are
overly confident.
These models were fine-tuned on our LetGo dataset for 20 epochs (shoes:
N=12,515; handbags: N=12,222). We opted not to use a learning rate schedule due
to the small amount of data.
Aesthetic quality baseline. We also considered a baseline aesthetic qual-
ity prediction task to test whether existing models that capture photographic
aesthetics can generalize to predict product image quality. We fine-tuned an
Inception v3 network on the AVA Dataset [Murray et al., 2012]. To transform
predictions into outputs suited to our dataset, we binned the mean aesthetic
annotation score from AVA into positive, neutral, and negative labels. The model
was fine-tuned on AVA for 5 epochs.
Evaluation. We used a binary “forced-choice” accuracy metric: starting from
a held-out set of positive and negative examples, we considered the network
output to be positive if x2>x0, effectively forcing the network to decide whether
the positive elements outweigh the negative ones, removing the neutral output.
By this metric, our best shoes model achieved 84.34% accuracy and our best
handbag model achieved 89.53%. This indicates that our model has a reasonable
chance of agreeing with crowd workers about the overall quality sentiment of a
listing image. (If we include neutral images and predictions and simply compare
the 3-way predicted output to the ground truth label on the entire evaluation
59
set, our handbag model achieves 64.09% top-1 accuracy and our shoes model
achieves 58.36%.) For comparison, the baseline aesthetic model achieved 68.8%
accuracy on shoes images and 78.8% accuracy on handbags. This shows that
product image quality cannot necessarily be predicted by aesthetic judgments
alone, and that our dataset constitutes a unique source of data for the study
of online marketplace images. Our image quality model gives us a black box
understanding of image quality in an uninterpretable way. Next, we investigate
what factors influence image quality.
3.5.1 What Makes a Good Product Photo?
Guiding sellers to upload good photos for their listings is a challenge that almost
all online eCommerce sites face. Many sites provide photo tips or guidelines for
sellers. For example, Google Merchant Center [Google, 2019]. suggests to “use a
solid white or transparent background”, or to “use an image that shows a clear
view of the main product being sold”. eBay also provides “tips for taking photos
that sell” [eBay, 2019], including “tip #1: use a plain, uncluttered backdrop to
make your items stand out”, “tip #2: turn off the flash and use soft, diffused
lighting”. In addition, many online blogs and YouTube channels also provide
tutorials on how to take better product photos.
Despite the abundance of product photography tips, little previous work
has validated the effectiveness of these strategies computationally (with the
exception of [Chung et al., 2012, Goswami et al., 2011]). Although there is a
robust line of research on computational photo aesthetics (e.g., [Schwarz et al.,
2016]), product photography differs greatly in content and functionality from
60
other types of photography and is worth special examination.
In this work, we leverage our annotated dataset, and conducted the first
computational analysis of the impact of common product photography tips on
image quality. Unlike previous work [Chung et al., 2012, Goswami et al., 2011]
that analyzed the impact of image features on clicks, we evaluate directly on
potential buyers’ perception of image quality. In a later section, we then show
how image quality can in turn predict sales outcomes.
Selecting Image Features
In order to select the image features to validate, we took a qualitative approach
and analyzed 49 online product photography tutorials. We collected the tutorials
through Google search queries such as “product photography”, “how to take
shoe photo sell”, and took results from top two pages (filtering out ads). We man-
ually read and labeled the topics mentioned in these tutorials and summarized
the most frequently mentioned tips. Out of the 49 tutorials we analyzed, the most
frequent topics were: (1) Background (mentioned in 57% of the tutorials): key-
words included white, clean, uncluttered; (2) Lighting (57%): soft, good, bright;
(3) Angles (40%): multiple angles, front, back, top, bottom, details; (4) Context
(29%): in use; (5) Focus (22%): sharp, high resolution; (6) Post-production (22%):
white balance, lighting, exposure; and (7) Crop (14%): zoom, scale.
Based on the qualitative results, as well as referencing previous work [Chung
et al., 2012, Goswami et al., 2011], we defined and calculated a set of image
features to analyze for their impact on image quality. There are three types
of image features that we considered: (1) Global features such as brightness,
61
contrast, and dynamic range; (2) Object features based on our object detector
(more details on that below); and (3) Regional features focusing on background
and foreground. Table 3.1 contains the definition and example images for our
complete set of image features.
Calculating Image Features
Global image features can be computed without extra information, but object
and regional features require knowledge of the object that appears in the image.
We trained an object detector that could detect bounding boxes for our product
categories.
Object detection. The process for building shoe detectors and handbag
detectors is the same. First, we collected and manually verified a dataset of
170 shoe images from the ImageNet dataset [Deng et al., 2009]. Those images
were already labeled with the bounding box around each shoe, and they vary
across many visual styles and contexts (not just online marketplaces). We also
augmented the ImageNet images with our own from the online marketplace
mentioned in Sec. 3.3.2. We designed a crowdsourcing task and labeled 500
images from the online marketplaces image dataset. Crowd workers were asked
to draw the bounding box around each single shoe in the image, and each image
was assigned to two distinct crowd-workers in order to ensure quality labeling.
We filtered out labels where the overlap between two bounding boxes were less
than 50%. In total, we gathered 650 shoe images from both ImageNet and our
online marketplace datasets, with bounding boxes around each shoe.
Next, we trained our shoe detector by using the Tensorflow Object Detection
62
Feature Name Definition Low High
Global Features:
brightness 0.3R + 0.6G + 0.1B
contrast Michelson contrast
dynamic_range grayscale (max - min)
width the width of the photo in px
height the height of the photo in px
resolution width * height / 106
Object Features:
object_cnt # of objects detected
top_space bounding box top to top of im-age in px
bottom_space bounding box bottom to bot-tom of image
left_space bounding box left to left of im-age
right_space bounding box right to right ofimage
x_asymmetry abs(right_space - left_space)/width
y_asymmetry abs(top_space - bottom_space)/height
Regional Features: (fg: foreground; bg: background)
fgbg_area_ratio # pixels in fg / bg
bgfg_brightness_diff brightness of bg - fg
bgfg_contrast_diff contrast of bg - fg
bg_lightness RGB distance from a purewhite image
bg_nonuniformity standard deviation of bg pixelsin grayscale
Table 3.1: Image feature definitions and example images
63
API and Single Shot Multi-box detector method [Liu et al., 2016]. The training
set included 300 images randomly selected from our labeled datasets. Finally,
we evaluated the performance of our detector on a validation set of 350 images.
We achieved a mean Average Precision (mAP0.5) of 0.84 for the shoe detection.2
We repeated the same process for the handbag detector, and reached similar
performance.
The resulted object detectors output bounding boxes around the shoes and
handbags in the image, which we then used to compute object and regional
features. In particular, for regional features, we used GrabCut algorithm [Rother
et al., 2004] to segment the foreground and background, initializing GrabCut
with the detected bounding boxes as the foreground region. Then we computed
the lightness and non-uniformity of the background, as well as differences in
brightness and contrast between background and foreground. Table 3.1 contains
all details and example images for our image features.
Regression Analysis
After calculating the image features, we analyzed their impact on image quality
through multiple ordered logistic regression. Our dependent variable is the three
image quality labels (bad, neutral, good) annotated through the crowdsourcing
task in Sec. 3.4. We choose to use the image label rather than raw image quality
scores for the dependent variable because there is less noise in the label (as we
did majority voting to get image labels). All analysis was done on the LetGo
image dataset, as that’s the set of images we collected manual image quality label
on.
2mAP0.5 means an image counts as positive if the detected and groundtruth bounding boxes
overlap with an intersection-over-union score greater than 0.5.
64
Detection: shoes. We first report on analysis of the shoe images. In our
dataset, 7% of the images did not have a target object detected, 22% has one target
object detected, 70% had two or more targeted objects detected. A chi-squared
test showed that there were significant differences in the distribution of image
quality labels across different number of target objects detected (χ2 = 463.68,
p < .001). Having at least one target object detected makes the image 2.7 times
more likely to be labeled as “good” quality.
Detection: handbags. Similarly, for handbags, 8% of our images did not
have a target object detected (with the detection score threshold of .90). A chi-
squared test showed that there was a significant difference in the distribution of
image quality labels between having a target object detected and not (χ2 = 60.34,
p < .001). Having the target object detected makes the image 1.4 times more
likely to be labeled as “good” quality.
To ensure the regression analysis remained accurate, we manually verified
detection accuracy on a subset of 2,000 images of shoes and 2,000 images of
handbags. We then conducted ordered logistic regression on this manually-
verified subset using image features as independent variables to predict the
image quality label. Results are shown in Table 3.23.
Brightness/background. On a high level, we confirmed brighter images are
more likely to be labeled as high quality for both product categories. In addition,
the non-uniformity of the background makes it less likely for an image to be
labeled as high quality. These two features coincide with the most commonly
mentioned product photography tips, background and lighting.
3Note the regression results do not differ significantly when we include the entire dataset,
showing that features extracted from automatic object detection are robust.
65
Shoes Handbags
Feature Name Estimate SE Estimate SE
brightness 3.30∗∗∗ (.96) 3.46∗∗∗ (.92)
contrast 1.79 (1.26) 4.89∗∗∗ (1.44)
dynamic_range 2.22∗ (1.104) .47 (1.67)
resolution -.10 (.06) -.21 (.19)
x_asymmetry -0.54 (.68) -2.26∗∗ (.84)
y_asymmetry -0.74 (.49) .73 (.46)
fgbg_area_ratio -.35∗∗∗ (.06) -.17∗∗∗ (.03)
bgfg_brightness_diff .59 (.48) -.11 (.53)
bgfg_contrast_diff -.52 (.54) -.80 (.42)
bg_lightness -.46 (.56) .16 (.55)
bg_nonuniformity -1.6∗ (.633) -4.84∗∗∗ (.64)
0|1 3.59∗∗ (1.20) 4.64∗∗∗ (1.22)
1|2 5.46∗∗∗ (1.21) 6.13∗∗∗ (1.22)
AIC: 4170.76 4169.92
Significance codes: ∗p < .05 , ∗∗p < .01 , ∗∗∗p < .001
Table 3.2: Ordered logistic regression coefficients predicting image quality
labels
Crop/zoom. We found mixed evidence around the crop of the images. For
both product categories, higher foreground to background ratio makes it less
likely for an image to be labeled as high quality, suggesting that the product
should be properly framed and not too zoomed-in.
Symmetry. Interestingly, we found that for shoes, asymmetry does not signif-
icantly contribute to perception of quality, but for handbags, horizontal asym-
metry moderately contributes to a lower perception of quality. One potential
explanation for this difference could be that shoes and handbags have different
product geometry dimensions (tall v.s. wide), and sellers would take pictures
with different orientations resulting in different distribution in vertical and hori-
zontal asymmetry to begin with. Indeed, we observed that handbag images were
slightly more likely to be in portrait (width<height) orientation than landscape
orientation (24% v.s. 22%, χ2=28.6, p<.001). Further investigation is necessary to
66
understand how symmetry impact the perceived quality of product photos.
Contrast. Finally, we observed that the difference in contrast between back-
ground and foreground can impact perceived image quality. In other words, a
good quality product image’s background should have low contrast, and the
foreground (the product) should have high contrast. The difference in brightness
between background and foreground, or the lightness of the background were
not significant in making an image more likely to be labeled as high quality —
suggesting a uniform background, either dark or bright, could be both effective,
and potentially work better for different colored products.
3.6 Marketplace Outcomes
In the previous section, we have shown that our trained models can automatically
classify user-generated product photos with an accuracy of 87% (averaged across
two product categories, shoes and handbags). Now leveraging the quality
scores predicted using our models, we proceed to examine how image quality
contributes to real-world and hypothetical marketplace outcomes.
We focus on two complementary marketplace outcomes: (1) Sales: whether an
individual listing with higher quality photos is more likely to generate sales; and
(2) Perceived trustworthiness: whether a marketplace with higher quality photos is
perceived as more trustworthy.
The first outcome, sales, is naturally important for online marketplaces. As
most of these platforms operate on a fee-based model (i.e., the platform charges
a flat or percentage fee when an item is sold), sales are directly linked to the
67
revenue and success of the marketplaces. Further, whether an item is sold is also
likely associated with higher user satisfaction.
Second, the perception of whether a platform is trustworthy is important for
the platform’s initial adoption and growth [Choi and Mai, 2018]. Previous work
has shown that the perceived trustworthiness of online marketplaces has a strong
influence on loyalty and purchase intentions of consumers [Hong and Cho, 2011],
retention [Sun, 2010], as well as creating a price premium for the seller [Pavlou
and Dimoka, 2006]. Several factors of the website’s visual design are known to
impact trust (e.g. complexity and colorfulness [Reinecke et al., 2013]), but the
effect of product image quality on trustworthiness remains an open question.
3.6.1 Image Quality and Sales
For the first outcome (sales of individual listings), we used log data from eBay
for analysis. Since merchants sometimes sell many quantities of the same item,
we predict whether a listing sold at least once before it expired. To that end, a
sample of balanced sold and unsold listings was created for the two product
categories we studied — shoes and handbags — see details of data sampling in
Sec. 3.3.2.
We first used logistic regression to understand how image quality relates to
trust. We considered three different models: (1) a baseline model using metadata
information about the listings, including the number of days the listing has been
on the platform, listing view count, and the item price; (2) a model including
image quality prediction score (the log-probability that an image will be classified
as high quality by models trained on our dataset), in addition to baseline features;
68
Shoes (N=130K) Handbags (N=32K)
(1) (2) (3) (1) (2) (3)
(Intercept) −1.09∗∗∗ −1.14∗∗∗ −1.11∗∗∗ −2.66∗∗∗ −2.61∗∗∗ −2.42∗∗∗
(0.05) (0.05) (0.05) (0.07) (0.07) (0.07)
# Days (Log) 0.26∗∗∗ 0.26∗∗∗ 0.26∗∗∗ 0.64∗∗∗ 0.63∗∗∗ 0.60∗∗∗
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
# Views (Log) 1.07∗∗∗ 1.08∗∗∗ 1.08∗∗∗ 0.94∗∗∗ 0.95∗∗∗ 0.98∗∗∗
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
Price (Log) −0.57∗∗∗ −0.56∗∗∗ −0.57∗∗∗ −0.31∗∗∗ −0.32∗∗∗ −0.36∗∗∗
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
Img. Quality 0.16∗∗∗ 0.15∗∗∗ 0.22∗∗∗ 0.15∗∗∗
(0.01) (0.01) (0.01) (0.01)
Img. Aesthetic 0.08∗∗∗ 0.31∗∗∗
(0.01) (0.02)
AIC 116,968 116,525 116,414 30,685 30,453 30,026
Note: ∗p<0.05; ∗∗p<0.01; ∗∗∗p<0.001
Table 3.3: Image quality predicted by our models is positively associated
with higher likelihood that an item is sold (1.17x more for shoes,
and 1.25x more for handbags).
and (3) a model including both image quality and aesthetic quality scores (trained
on the AVA dataset), in addition to baseline features. The results of regressions
for both product categories predicting whether an item is sold are reported
in Table 3.3.
From the regression analysis, we show that image quality predicted by our
models is associated with higher likelihood that an item is sold (odds ratio 1.17
for shoes, 1.25 for handbags, p<.001). These results hold even when controlling
for the predicted aesthetic quality of images. Interestingly, both predicted image
quality and aesthetic quality are more strongly associated with the sales of
handbags than shoes (β=0.22 v.s. 0.16 for image quality, and β=0.31 v.s. 0.08 for
handbags), potentially signaling that handbag is a more visual product category
than shoe.
69
Figure 3.4: Hypothetical marketplace mock-ups used for our user exper-
iment. From left to right, showing images with high quality
score, low quality score, and stock imagery.
However, in terms of model performance, both image quality and aesthetic
quality only resulted in small improvements to the baseline model. We illustrate
the model performance through prediction accuracy — 10-fold cross validation
showed that the baseline model can predict whether an item will be sold with an
accuracy of 80.3% for shoes, and 74.7% for handbags. Both image quality and
aesthetic quality improved the prediction accuracy only marginally (around 1%).
These findings suggest that both image quality and aesthetic quality are associ-
ated with higher likelihood of sales for individual listings on online marketplaces,
though both have limited power in improving the accuracy of sales prediction.
Future work could explore the difference among different product categories,
or the relationship between image quality and other metrics for online market-
places, such as “sellability” [Liu et al., 2017], and click-through-rate [Goswami
et al., 2011, Chung et al., 2012].
3.6.2 Perceived Trustworthiness
For the second outcome, perceived trustworthiness of the marketplace, we
designed a user experiment to compare the effects of good/bad quality user-
70
generated images compared to stock imagery. Our goal here is to show the
potential application of our image quality models in improving the perceived
trustworthiness of online marketplaces. Our hypothesis is that high quality
marketplace images (as selected by our models) will lead to the highest per-
ception of trust, followed by stock imagery, and then low quality marketplace
images will lead to the lowest perception of trust. This hypothesis is rooted in
the fact that uncertainty and information asymmetry are fundamental problems
in online marketplaces that limit trust [Akerlof, 1978]. Stock images, though high
in aesthetic quality, do not help reduce the uncertainty of the actual conditions
of the product being sold. High quality user-generated images could help bridge
the gap of information asymmetry and reduce uncertainty, therefore increasing
the trust.
To test this hypothesis, we built three hypothetical marketplaces mimicking
the features of popular online platforms (see Figure 3.4), each populated with
(1) high quality marketplace images selected by our model, (2) low quality
marketplace images selected by our model, and (3) stock imagery from the UT
Zappos50K dataset [Yu and Grauman, 2014, Yu and Grauman, 2017]. Specifically,
we designed a between-subject study with three conditions, varying the images
of the listings (good; bad; and stock). Each participant was randomly assigned
to one condition and saw three example mock-ups of a hypothetical online peer-
to-peer marketplace website, each populated with 12 images randomly drawn
from a set of 600 candidate images.
We prepared the candidate images for each condition in the following ways.
All of our images were from the shoes category, but could easily be expanded to
other categories. For the “good” condition, the candidate images were 600 images
71
randomly drawn from the top 5% LetGo images as predicted by our image quality
model (to have high image quality). Similarly, candidate images used for the
“bad” condition were drawn from the bottom 5% of LetGo images predicted
by our image quality model. For the “stock” image condition, we randomly
sampled 600 images from the UT Zappos50K [Yu and Grauman, 2014, Yu and
Grauman, 2017] as candidate images.
The main dependent variable for this experiment is the perceived trustwor-
thiness of the marketplace, which could be measured in a few different aspects.
Therefore, we developed a six-item trust scale based on adaptations of previous
work on trust in online marketplaces [Corbitt et al., 2003], shown in Table 3.4.
Each participant was requested to rate the marketplace they were shown on a
5-point Likert scale. We take the average of participant’s responses to all items in
the scale as the “trust in marketplace” score.
Results
The experiment was pre-approved by Cornell’s Institutional Review Board,
under protocol #1805007979. We issued the task through Amazon Mechanical
Turk and recruited 333 participants, paying 50 cents per task.
We retained 303 submissions after initial filtering, evenly distributed across
three conditions. We filtered out the submissions that were completed too fast or
too slow (trimming the top and bottom 5% based on task completion time), as
previous work has shown that filtering based on completion velocity improves
the quality of submissions [Ma et al., 2017b, Wilber et al., 2017].
Overall, participants reported highest level in marketplaces populated with
72
good images, followed by stock imagery, and the lowest level of trust in market-
places populated with bad images (p<.001). The average perceived trustworthi-
ness of marketplaces per condition is shown in Figure 3.1.
The “trust gap” between high quality user-generated images and stock im-
agery suggests that high “quality” images on online marketplaces do not neces-
sarily have to be more aesthetically pleasing. Our finding corroborates previous
findings that show users prefer actual images over stock imagery because they
give an accurate depiction of what the product looks like [Bland et al., 2007].
Stock images could be perceived impersonal or “too good to be true” in on online
peer-to-peer setting. High quality user-generated images, on the other hand,
reduce uncertainty and information asymmetry in online transaction settings,
therefore increasing trust.
Taken together, our marketplace experiments showed that our image quality
models could effectively pick out images that are of high quality and can increase
the perceived trustworthiness of online peer-to-peer marketplaces, even out-
performing stock imagery. The results of the experiment also suggest potential
real-world applications of our image quality dataset as well as prediction models,
by automatically ranking, filtering and selecting high quality images to present
to the users to elicit feelings of trust.
3.7 Discussion and Conclusion
This work presents a computational understanding of user-generated images
in online marketplaces and how image quality relates to trust. By gathering
and annotating a large-scale dataset of user-generated product images from
73
Item Definition
General How trustworthy do you think this marketplace is?
Technical I believe that the chance of having a technical failure on this
marketplace is quite small.
Risk I believe that online purchases from these sellers are risky.
Expectation I believe that products from these sellers will meet my expecta-
tions when delivered.
Care I believe that these sellers care about its customers.
Fidelity I believe that the photos accurately represent the condition of
the products.
Note: Response to each item is based on a 5-point Likert scale (strongly disagree
to strongly agree)
Table 3.4: Marketplace perceived trustworthiness scale
online marketplaces, we were able to develop a deeper understanding of the
visual factors that improve image quality, while reaching a decent accuracy
(≈87%) for predicting image quality. In particular, high quality user-generated
images selected by our model outperform stock imagery in eliciting perceived
trustworthiness in the platform by a potential user.
These findings have important implications for the future of increasingly
mobile-based online marketplaces; the dataset can also be useful for the broader
computer vision community, e.g., providing more examples of real-world images
for domain adaptation tasks. One can also leverage this work to build tools to
help users take better product photos, potentially through novel interfaces such
as Augmented Reality. These tools can potentially facilitate trust between sellers
and buyers on online marketplaces.
The fact that high quality user-generated images outperform stock images
highlights how the dynamics between images and trust have evolved over the
years. User-generated images may continue to play a more important role, as
users believe that they reduce more uncertainty and are more useful in expecta-
74
tion setting. At the same time, low quality user-generated images perform badly
in generating perceived trustworthiness. Future platforms can consider algorith-
mically filtering out low-quality user-generated images, or provide suggestions
on which images to feature for users to promote overall trust on the platform.
This work is also not without limitations. The dataset, while large in size, only
covered two product categories. Predicted image quality also only has limited
prediction power in whether an item is sold. Future work could further enrich
the dataset, expanding to other categories, such as images of Airbnb listings.
Future work can also explore how image quality might indirectly influence sales
(e.g., through increased view count).
Finally, from a networked trust perspective, the work presented in this chapter
furthers our understanding of the first aspect of networked trust (CMC). By
comparing user-generated images with stock images, I show that trust online
through images is not simply established with aesthetics or photographic quality,
but rather, the principles of uncertainty reduction. While photographic quality
definitely plays a role, and low-quality user-generated images perform badly in
eliciting perceived trustworthiness, there might be a “uncanny valley” when it
comes to image aesthetics and trust. When something appears “too good to be
true”, such in the case of stock images, user trust might be negatively impacted.
High-quality user-generated images, on the other hand, may have hit a sweet
spot between aesthetic quality and expectation management and uncertainty
reduction, thus gaining most trust.
Importantly, it is important to point out that it is a conscious choice to not focus
on image cues of people but rather image cues of products in this research inquiry
on the relationships between images and trust, as prior research has pointed out
75
racial bias in state-of-art facial recognition algorithms [Buolamwini and Gebru,
2018]. However, images of people are perhaps even more important in the context
of interpersonal trust and trust in media and information. Future work can
further investigate how image cues of people affect trust and its interaction with
bias. For example, does images of faces influence how politicians are trusted? Is
it possible to study such mechanisms computationally while accounting for racial
and gender bias in current facial recognition systems? Finally, as computer vision
techniques continue to advance, images and videos can both be manipulated
to distort the true representation of people. For example, a doctored video of
House Speaker Nancy Pelosi reached millions of viewers on Facebook [Rini,
2019]. Understanding how such manipulated image and video cues can impact
interpersonal trust remains important future work for networked trust, especially
in the context of misinformation.
76
CHAPTER 4
LANGUAGE OF TRUST: SELF-PRESENTATION AND PERCEIVED
TRUSTWORTHINESS OF AIRBNB HOST PROFILES
4.1 Introduction
The last chapter presented a study of trust in one of the more traditional types
of online exchange platforms, online commerce platforms, where I showed that
image cues are important to establish trust. In this chapter, we examine how
language cues establish trust in a relatively new type of social exchange platform
— sharing economy.
I view sharing economy platforms among the third wave of the digitaliza-
tion of exchange — the digitalization of the exchange of resources. Challenges
about trust in this wave of digitalization of exchange center around the fact
that people need to have high levels of trust in strangers. Not only people need
to trust strangers with their financial resources, personal information, but also
physical safety. Language has been shown to be especially important in the early
stages of relationship development among strangers, through the process of
self-disclosure [Cozby, 1972]. As sharing economy platforms gain popularity, it
is important to understand how language cues establish interpersonal trust in
this context.
In particular, this chapter examines how language cues establish trust on
Airbnb through self-disclosure in host profiles. 1 I discuss trust on Airbnb in a
1The work presented in this chapter was published at CSCW 2017 as Self-disclosure and
Perceived Trustworthiness of Airbnb Host Profiles, and in ICWSM 2017 as A Computational Approach
to Perceived Trustworthiness of Airbnb Host Profiles [Ma et al., 2017a, Ma et al., 2017c].
77
dyadic context, and measure perceived trustworthiness of host using surveys.
Airbnb is an online lodging marketplace for short-term peer-to-peer rentals,
facilitating monetary and social exchange between individuals [Lampinen and
Cheshire, 2016]. On Airbnb, hosts can list places (e.g. rooms, apartments, houses,
or even boats and castles) for guests to rent. The guest is often a temporary visitor,
and is not acquainted with the host beyond Airbnb. As of July 2019, Airbnb
reports six million listings, and 500 million guest arrivals all-time through the
platform2.
The main utility of Airbnb — identifying potential lodging resources offered
by unknown individuals — comes with risks that affect both guests and hosts
who wish to participate in the exchange. A potential host may worry about
guests damaging their property. A potential guest may fret about their physical
safety, the truthfulness of the quality of the property being advertised, or whether
the host would be kind enough to provide assistance in exigencies [Ert et al.,
2016]. Establishing guest-host trust helps manage such uncertainties and risks —
making trust a crucial factor for the success of such social exchange sites.
There are several ways that Airbnb designs for trust. Airbnb has an assur-
ance policy and a reputation system in place, in addition to making information
about the host and property readily available before booking. On Airbnb, each
host has a profile page that includes photos, a text-based self-description, social
media verification status, and reviews (if any) from other Airbnb users who
have stayed with the host. These profiles contribute to a guest’s decision-making
process [Newman and Antin, 2016], and help establish perceived trustworthi-
ness [Ert et al., 2016]. In this work, we focus on host profiles, especially the
2https://press.airbnb.com/fast-facts/
78
text-based self-description and its role in establishing the perceived trustworthi-
ness of hosts in the eyes of potential guests.
There are two steps towards the understanding of how language cues estab-
lish trust on Airbnb. In the first step, we examine how hosts describe themselves
on their Airbnb profile pages. We use a mixed-methods study to develop a
categorization of the topics that hosts self-disclose in their profile descriptions,
and show that these topics differ depending on the type of guest engagement
expected. We also examine the perceived trustworthiness of profiles using
topic-coded profiles from 1,200 hosts, showing that longer self-descriptions are
perceived to be more trustworthy. Further, we show that there are common
strategies (a mix of topics) hosts use in self-disclosure, and that these strategies
cause differences in perceived trustworthiness scores. We then show that the per-
ceived trustworthiness score is a significant predictor of host choice — especially
for shorter profiles that show more variation. The results are consistent with
uncertainty reduction theory, reflect on the assertions of signaling theory, and
have important design implications for sharing economy platforms, especially
those facilitating online-to-offline social exchange.
In the second step, we developed a novel computational framework to pre-
dict the perceived trustworthiness of host profile texts. To achieve this goal, we
extended the previous dataset from 1,200 to 4,180 Airbnb host profiles anno-
tated with perceived trustworthiness. To the best of our knowledge, the dataset
along with our models allow for the first computational evaluation of perceived
trustworthiness of textual profiles, which are ubiquitous in online peer-to-peer
marketplaces. We provide insights into the linguistic factors that contribute to
higher and lower perceived trustworthiness for profiles of different lengths.
79
4.2 Related Work
User profiles are an important part of many online systems, and serve a variety of
functions [Uski and Lampinen, 2014]. In social networking sites, profiles provide
an identity for the user that persists over time and the myriad of interactions
on the site [boyd and Ellison, 2007]. In online dating sites, profiles provide
self-disclosure that attracts the interest of other users, while limiting the risks
associated with outright deception [Gibbs et al., 2010, Toma et al., 2008].
In each of these contexts — especially in services that lead to offline interac-
tions — a key function of the profile is to establish perceived trustworthiness [Ert
et al., 2016, Guttentag, 2015]. Here, we distinguish between trustworthiness and
trust. Trustworthiness is an attribute of a trustee [Hardin, 2002, Kiyonari et al.,
2006], while trust is exhibited by a trustor (e.g. a decision to take risks in an
economic game [Berg et al., 1995, Cook et al., 2005]). As our focus is on the host,
the trustee, we focus on perceived trustworthiness as an attribute of the host.
One approach to understanding how profiles are used to establish trustwor-
thiness is the Profile as Promise framework [Ellison and Hancock, 2013], which
uses the notion of a promise to characterize the function of a profile. In this
view, the profile is a psychological contract between the profile holder, and the
viewer that future interactions (e.g. with a date, a car driver, or an Airbnb host)
will take place with someone who does not differ fundamentally from the per-
son represented in the profile. The notion of a promise has been successfully
applied to various contexts that require good faith to operate, including online
dating [Ellison and Hancock, 2013] and job hunting [Rousseau and Greller, 1994].
The Profile as Promise perspective argues that the content and characteristics
80
of the disclosures, or promises, made in user profiles should be diagnostic
of trustworthiness perceptions. Within this framework we draw on theories
from communication, and from economics, to form specific expectations about
disclosures and their perceptions. For example, communication scholars have
used uncertainty reduction theory [Berger and Calabrese, 1975] to show that
strangers are concerned with increasing the predictability about the behavior of
both themselves and others in the interaction that occurs when they first meet.
Uncertainty reduction theory has been used in research on dating sites to explain
how much information may be shared in profiles [Gibbs et al., 2010], using self-
disclosure as the process of making the self known to others [Derlega et al., 1987].
According to this approach, people should disclose as much information in their
host profiles as they feel comfortable sharing — a directive that presents various
challenges, including the risk of over-exposure if the profile is public [Gibbs et al.,
2010, Bazarova and Choi, 2014]. Nevertheless, the more information disclosed in
a profile, the more likely it is perceived as trustworthy.
According to the Profile as Promise framework, one way to understand
what kinds of information people disclose in the kind of static and asymmetric
context of online profiles is signaling theory [Ellison and Hancock, 2013], which
considers the relationship between explicitly stated signals and the underlying
qualities they are likely to represent [Donath, 2007, Spence, 2002]. Spence’s
original application of signaling theory [Spence, 2002] explored how potential
employees in a labor market tried to convey attributes that cannot be observed
directly, such as reliability or goodness-of-fit with a company’s culture. Signaling
theory was also concerned with how such signals can be assessed, for example,
by the employer. Some signals, called conventional signals, are relatively easy
to fake; such as asserting that one is a reliable and hard worker, when in fact
81
one is not. Other signals, called assessment signals, are more difficult to fake;
such as claiming that one has a degree from a prestigious institution. Gambetta
used signaling theory to study how taxi drivers [Gambetta and Hamill, 2005]
and criminals [Gambetta, 2009] form impressions and assess the trustworthiness
of other parties based on cues, especially in fleeting initial interactions where
high uncertainty is involved. Here we examine how Airbnb hosts signal their
trustworthiness in their profile through self-disclosure and how these signals are
perceived.
4.3 Step 1: Self-Presentation and Perceived Trustworthiness of
Airbnb Host Profiles
Emerging literature is examining how people assess trustworthiness through self-
disclosures made in online profiles. The Profile as Promise [Ellison and Hancock,
2013] conceptual framework, for example, incorporates the risks and rewards
associated with assessing signals in a profile for whether said profile’s promises
can be trusted. Researchers had examined how individuals produce and assess
trustworthiness signals in online dating profiles [Toma et al., 2008] and in online
résumés [Guillory and Hancock, 2012]. However, we still know very little
about what people self-disclose, and how that information is evaluated for
trustworthiness in the context of sharing economy platforms such as Airbnb.
Given the importance of profiles, in this work we aim to advance our under-
standing of the type of content Airbnb hosts self-disclose in their profiles, and
to determine the impact of these disclosures on perceived trustworthiness and
host choice. We build on the Profile as Promise framework [Ellison and Han-
82
cock, 2013], drawing on theories from economics and communication to predict
what kinds of information hosts will disclose in their profiles, and what kinds
of disclosure will enhance trust. In particular, we apply uncertainty reduction
theory [Berger and Calabrese, 1975] to predict that both quantity and diversity
of information increases the perception of trust. We also draw on signaling
theory [Donath, 2007, Spence, 2002] to predict that specific kinds of information
can signal trustworthiness in a profile.
Specifically, we use a mixed-methods approach, with qualitative analysis,
large-scale annotation, and an online experiment to examine the text-based self-
descriptions of Airbnb host profiles. We qualitatively develop a categorization
scheme that characterizes the primary self-disclosure topics in these profiles. We
then quantitatively show how predictions from uncertainty reduction theory
and signaling theory apply to this case, revealing that an increase in the quan-
tity of content and the inclusion of specific topics can enhance perceptions of
trustworthiness. Finally, we use an online experiment to show that the perceived
trustworthiness of profiles is a significant predictor of host choice. Our results
have practical design implications for platforms facilitating social exchange in
the sharing economy.
Research Questions
In our examination of Airbnb profiles, our first objective, according to the Profile
as Promise framework, is to determine what kinds of information hosts provide
in their profiles to reduce uncertainty and signal trustworthiness.
RQ 1. What kinds of information do hosts self-disclose to signal their trustworthi-
83
ness?
The second question is concerned with how the information hosts disclose
in their Airbnb profiles translates into perceived trustworthiness by guests. On
Airbnb, like other peer-based sharing economy services, signals of trustworthi-
ness are particularly important as trust is critical for social exchange [Cheshire
and Cook, 2004,Lampinen and Cheshire, 2016]. Trust on Airbnb and other online
marketplaces is tied to ratings and reputation. However, ratings on Airbnb tend
to not be informative as they are skewed high, and there is initial evidence that
the number of reviews received is predictive of room sales even when controlling
for scores [Lee et al., 2015a, Zervas et al., 2015]. Research suggests that profile
information matters on Airbnb: in one study, profile images were linked to the
perceived trustworthiness of hosts and higher prices [Ert et al., 2016]. At the
same time, the study showed that online review scores had no effect on the listing
price, although profile text was not considered. We therefore ask:
RQ 2. What is the effect of different types of self-disclosure on perceived trustworthi-
ness?
Note that it is not immediately clear that trustworthiness maps directly to
the choice of host. Choice can clearly be influenced by other factors, such as
assurance [Cheshire, 2011]. There is initial data-driven evidence that visual-
based trustworthiness impacts choice [Ert et al., 2016], even when reputation
scores are manipulated to increase their variance.
Given the difficulty in establishing the causal link between profile disclosures
and guest decision-making, we conduct an experiment that isolates the trustwor-
thiness of a profile’s disclosures from other external factors — such as reputation
84
indicators — and manipulates the effect of low versus high trustworthiness
profiles on a decision-making task. In this experiment, we focus on addressing
the following research question:
RQ 3. Do profile-based perceptions of trustworthiness predict choice of host on
Airbnb?
4.3.1 Study 1 — How do Hosts Self-Disclose?
The primary goal of Study 1 is to uncover what Airbnb hosts self-disclose in the
text-based self-descriptions in the profile. To the best of our knowledge, there is
no established coding scheme for self-disclosure in this particular context. For
this reason, Study 1 uses qualitative methods to develop, validate and apply a
coding scheme for self-disclosure in Airbnb host profiles. We accomplished this
task using a two-phase approach. In phase one, we developed and validated a
coding scheme for topics of self-disclosure by qualitatively analyzing Airbnb
profiles and using an inductive and iterative approach to identify categories.
In phase two, we applied the coding scheme to a large set of host profiles, and
examined patterns of self-disclosure on Airbnb.
Phase 1 — Developing and Validating Coding Scheme
Step 1: Development
To create the self-disclosure coding scheme for Airbnb, we used an iterative,
inductive analysis for content-topic categories of information in host profiles. As
this study is exploratory in nature, we established some guidelines for developing
85
the initial coding scheme. In particular, when creating an Airbnb profile page, the
website prompts the host to share a few details about themselves, calling out three
types of self-disclosure: “things you like”, “style of traveling or hosting”, and
“life motto”. We were cognizant of the Airbnb interface prompt and used it as a
starting point, fitting codes to the prompt topics and refining them according to
the content, but not restricting our coding to topics suggested by these prompts.
For this step, we constructed a Development Dataset consisting of 300 sentences
randomly drawn from a weighted sample of 203 host profiles from 12 major
U.S. cities. The profiles were extracted from an open-sourced Airbnb dataset
collated by an independent organization, Inside Airbnb [Inside Airbnb., 2016].
Non-English profiles were filtered out. We provide more details of the full dataset
below.
Two raters independently coded the topics in each of the 300 sentences in the
Development Dataset, using the qualitative data analysis and research software
Atlas.ti. In addition to Airbnb prompts, the coders also considered topics used in
previous self-disclosure studies [Jourard and Lasakow, 1958,Ma et al., 2016,Rubin,
1975]. After a full round of independent coding, the two coders compared their
codes and deliberated, further merging the codes into concepts and topics. This
analysis and coding process resulted in nine initial topic categories.
Step 2: Adjustment and Validation
In this step, we evaluated, adjusted, and validated the coding scheme for
reliability and coverage, i.e. the percentage of sentences our codes applied to.
In order to apply the coding scheme to a large set of host profiles on Airbnb,
we designed a web interface to recruit annotators from Amazon Mechanical
86
Topic Agreement Description
Vote- Vote- R1-
R1 R2 R2
Interests & Tastes .77 .85 .78 Favorite books, music, hobbies, how I
spend weekends and evenings, favorite
ways of spending spare time.
Life Motto & Values .51 .56 .52 Life motto, values, philosophies; e.g. “Live
courageously, love passionately”.
Work or Education .83 .86 .79 Current or past job, school, major; e.g.
“I’m an architect and designer”.
Relationships .69 .62 .60 Family, significant other, pet; e.g. “I have
a beautiful 16 year old daughter, a little
sweet terrier Nora, two fish & a frog.”
Personality .80 .70 .65 e.g. “I am extremely down to earth and I
am a self-diagnosed work-a-holic”.
Origin or Residence .78 .69 .76 Where from, current residence, history of
moving; e.g. “I lived in D.C. for 5 years
and Philly for 2 years”; “We both really
love how much Chicago has to offer.”
Travel .78 .72 .83 Love for travel; past travels; favorite travel
destinations.
Hospitality .73 .54 .66 Welcoming or greeting, reasons for host-
ing, demonstrating availability; e.g. “We’re
delighted to be your hosts and tour advi-
sors during your stay here.”
Table 4.1: Topics of self-disclosure in Airbnb host profiles.
Turk (AMT). The web interface presented each individual sentence from the host
profiles, together with the initial topics and descriptions (some with examples)
that were developed through the coding process in the foregoing step. The
annotator was instructed to tag all topics that appeared in the sentence (a sentence
could mention multiple topics). If none of the topics applied, the annotator was
instructed to choose “other”.
We validated the reliability of the coding scheme using two metrics: the
level of agreement among crowd workers, and the level of agreement between
the crowd workers’ consensus and expert annotations, i.e. researchers from our
87
team.
To compute the first metric, the level of agreement among crowd workers, we
constructed the Initial Validation Dataset, consisting of 300 sentences drawn from
a new sample of 203 profiles. Sentences in the Initial Validation Dataset did not
overlap with those in the Development Dataset. We recruited crowd workers from
AMT to annotate sentences in the Initial Validation Dataset using a web interface
that we developed (paying $.02 per annotation), and computed Fleiss’ kappa
among the workers. There were four topics that had a Fleiss’ kappa score lower
than 0.5, indicating an unsatisfactory level of internal agreement. We iterated
on the initially developed set of topics to address this issue, adjusting the name
and description of two topics, and merged two closely related topics into one
(“Hosting Attitude” and “Hosting Action” to “Hospitality”). After the edits to
the coding scheme, eight topics remained, shown in the first column of Table 4.1.
To compute the second metric — the level of agreement between the con-
sensus from crowd workers and expert annotations — we constructed the Final
Validation Dataset, consisting of all 871 sentences from the text of a new batch of
203 profiles. The new profiles did not overlap with either those in the Development
Dataset or those in the Initial Validation Dataset. Again, we asked three crowd
workers to annotate each sentence, and used a majority voting rule to produce
the final vote across three workers. A topic label for a sentence was retained only
if at least two out of three workers indicated that the sentence mentioned that
topic.
In terms of coverage, in total, at least two voters agreed on one or more topics
for 91.5% of the 871 sentences (if we consider the workers’ using an optional
“other” category, a majority vote was achieved for 97.4% of the sentences). The
88
raters inspected the sentences where the workers did not reach an agreement,
and verified that they did not contain significant missed themes.
Finally, two raters coded a 300-sentence sample from the Final Validation
Dataset. We computed the agreement of these three different sources by calcu-
lating the pairwise Cohen’s kappa scores for the worker majority vote (vote in
Table 4.1), Rater 1 (R1), and Rater 2 (R2). The results of each pairwise agreement
computation are shown in Table 4.1. The results suggest moderate to almost
perfect agreement across all topics, and indicate that the coding scheme (the topic
names, descriptions, and the set-up of majority votes from AMT) is reliable.
Phase 2 — Applying the Coding Scheme to Profiles
With the coding scheme validated, we could now annotate a large set of host
profiles and examine the trends of self-disclosure. What might we expect for the
disclosures? According to the Profile as Promise framework [Ellison and Han-
cock, 2013], hosts should disclose information they believe will signal to potential
guests that they will be a trustworthy host. These disclosure goals should cause
hosts to prioritize the disclosure of information that enhances trustworthiness.
Signaling theory further suggests that perceptions of trustworthiness may be
affected by the kind of signal [Donath, 2007, Spence, 2002]. If this is the case,
then hosts should disclose more assessment signals (i.e. disclosures that can be
verified):
H1.1 Hosts will disclose more about categories that have more assessment value,
including Work or Education, and Origin or Residence, than about categories that have
more conventional value, including Interests & Tastes, and Personality.
89
The Profile as Promise framework also suggests that information should be
disclosed about the most relevant underlying qualities that the host is promising
to potential guests. In this context, the type of hosting situation, on-site versus
remote, should lead to different disclosure patterns. The on-site hosts (who share
their space with guests during their stay) need to signal what kind of person a
guest might meet. The remote hosts (who are not present) need to signal that the
guests will be taken care of in their absence. Previous work on Airbnb revealed
that on-site versus remote hosting is an important part of sociability within the
host-guest relationship [Ikkala and Lampinen, 2015, Lampinen and Cheshire,
2016]. When hosting on-site, guests and hosts may have more substantial face-to-
face interaction.
Given the increased likelihood of social interaction for on-site hosts, there is
uncertainty about whether the guests and hosts will get along. We can draw on
uncertainty reduction theory [Berger and Calabrese, 1975] to predict that on-site
hosts will disclose more information than remote hosts in an effort to reduce
the uncertainty for potential guests given that guests and hosts will socially
interact. In particular, on-site hosts should disclose more information relevant to
relationship development, such as one’s preferences and personality.
H1.2 On-site hosts will disclose more, especially for topics that can reduce uncertainty
during the interaction of sharing spaces, such as Interests & Tastes, and Personality,
than remote hosts.
We first report on the dataset of Airbnb profiles we used for this analysis
and throughout the rest of this work. Then, we describe the process of applying
the coding scheme to annotate a larger portion of the host profiles. Finally, we
discuss the results of testing the hypotheses using the annotated data.
90
Airbnb Dataset
To apply the coding scheme to a larger portion of Airbnb profiles, we used
the large-scale dataset collected by Inside Airbnb [Inside Airbnb., 2016]. Inside
Airbnb periodically scrapes the Airbnb website, making snapshots of Airbnb
listings from 35 cities in 13 countries (at the time of writing) available for down-
load. For each city, Inside Airbnb conducted a URL query through Airbnb search
and scraped all public listings. We manually examined 10 samples from five
cities in the dataset, visiting the Airbnb website for each entry to verify that
the scraped data is consistent with the actual listing. Since each listing is al-
ways associated with a host, the free text portion of the host profile is available
from the Inside Airbnb dataset. Some other metadata about hosts, in addition
to the host self-description (the focus of the present work) include: host ID (a
unique identifier for a host across the Airbnb platform), first name, type of listing
(Entire Home/Apt, Private Room, or Shared Room), and whether the host is a
“superhost” on Airbnb.
We limited our scope of analysis to U.S. and English-language host profiles
only. Host profiles from other countries may contain non-English phrases or char-
acters, introducing sources of noise, and making it difficult for crowd workers to
annotate. We performed source language detection using the Google Translate
API [Google, 2016] for each sentence in a host profile, and filtered out those
containing non-English sentences.
The Inside Airbnb data included 93,361 listings across 15 U.S. cities. We first
de-duplicated hosts from multiple listings by host ID. We used data from the 12
largest cities, excluding 3 cities with fewer than 1,000 unique hosts (Asheville,
Oakland, and Santa Cruz County). We verified that the exclusion of these three
91
cities did not affect the results from Study 1. For the remaining 12 cities, we
deduplicated 89,965 listings to obtain 67,465 unique hosts. Out of these unique
hosts, we further filtered out 20,710 (or 30.7% of the de-duplicated quantity)
host profiles with empty self-descriptions, and 6,750 (10.0% of the de-duplicated
quantity) that contained non-English phrases.
In the end, we had 40,005 non-empty, English-only unique host profiles
from 12 U.S. cities, with the following breakdown: New York (data collected
in Sep 2015; 14,513), Los Angeles (Jan 2016; 8,062), San Francisco (Nov 2015;
3,400), Austin (Nov 2015; 2,477), Chicago (Oct 2015; 2,149), Seattle (Jan 2016;
1,798), Washington D.C. (Oct 2015; 1,633), San Diego (Jun 2015; 1,522), Portland
(Sep 2015; 1,415), New Orleans (Sep 2015; 1,173), Boston (Oct 2015; 922), and
Nashville (Oct 2015; 941).
With our previously validated coding scheme, we annotated the topics of a
larger portion of host profiles from the above-mentioned Airbnb dataset using
AMT, following the exact same procedure as described for annotating the Final
Validation Dataset. We constructed the Annotation Dataset, consisting of all 4,377
sentences from 1,031 profiles, randomly selected using a weighted sample ac-
cording to the number of unique non-empty host profiles in each city. As the
coding scheme was the same as that used for the Final Validation Dataset, we
merged the results from the Annotation Dataset and the Final Validation Dataset ,
forming the Experiment Dataset to boost the amount of annotated data, resulting
in 5,248 annotated sentences from 1,234 profiles.3
Self-Disclosure Trends
3The Experiment Dataset and other data used in this work are available from https://
github.com/sTechLab/AirbnbHosts.
92
What do Airbnb hosts self-disclose in their profiles? We found that hosts were
most likely to talk about Origin or Residence (68.8%), followed by Work or Education
(60.29%) and Interests & Tastes (57.78%). There was substantial travel-related
disclosure including writing about Travel (47.89%) and demonstrating Hospitality
(52.76%). The topics that were less commonly mentioned were Relationships
(27.88%), and Personality (26.58%). The topic that was least mentioned was Life
Motto & Values (7.86%).
This pattern of results is partially supportive of H1.1 and the prediction
from signaling theory that hosts would disclose more assessment signals than
conventional ones. Consistent with the hypothesis was the frequent disclosure of
assessment signals regarding Origin or Residence and Work or Education, and the
low levels of disclosure regarding Personality. The frequent disclosure of Interests
& Tastes, however, did not line up with the hypothesis. The analysis below on
disclosures by host type provides some insight: the high rate of disclosure of
Interests & Tastes was driven in part by on-site hosts, which may have been part
of an effort to reduce uncertainty for guests who would be meeting their hosts.
Self-Disclosure Trends by Host Type
Addressing H1.2, we compared the self-disclosure of hosts based on the
type of property they offered: on-site versus remote. As hypothesized, on-
site hosts (M=66.12, SD=59.66) on average wrote longer profiles (measured by
word count) compared to remote hosts (M=55.85, SD=52.09), t(880)=3.07, p < .01.
Second, in terms of topics, we found that on-site hosts were more likely than
remote hosts to write about topics that signal their personality and tastes. We
calculated the percentage of profiles that mentioned each topic for different host
types, shown in Figure 4.1. For each topic of disclosure (identified by the first
93
0.8
0.6
Host Type
0.4 On−site
Remote
0.2
0.0
I L W R P O T H
.5 Topic of Self−Disclosure
Figure 4.1: Probability of disclosure per topic
2.5
2.0
1.5 Host Type
On−site
Remote
1.0
0.5
0.0
I L W R P O T H
.5 Topic of Self−Disclosure
Figure 4.2: Average number of sentences per topic
Figure 4.3: Self-disclosure trends by topic and host type. The error bars
represent one standard error.
94
Number of Sentences Likelihood of Disclosure
letter on the x-axis), we show the percentage of profiles of each host type that
mentioned that topic (y-axis). As the figure reveals, on-site hosts were more likely
to write about topics of Interests & Tastes (χ2=18.57, df=1, p < .001) and Personality
(χ2=29.18, df=1, p < .001); and less likely to mention Origin or Residence (χ2=4.17,
df=1, p < .05). Note that the results for Interests & Tastes and Personality remain
statistically significant with Bonferroni correction for multiple tests.
We also compared the number of sentences used for the different disclosure
topics, which shows similar trends (Figure 4.2). For Interests & Tastes, on-site
hosts on average wrote more sentences (M=1.87, SD=.08) than remote hosts
(M=1.59, SD=.05), t(535)=3.09, p < .01. For Personality, we also see that on-site
hosts on average wrote more sentences (M=1.42, SD=.06) than remote hosts
(M=1.17, SD=.03), t(260)=3.64, p < .001. Finally, on-site hosts on average wrote
more sentences mentioning Life Motto & Values (M=1.56, SD=.19), t(48)=2.34,
p < .05 than remote hosts (M=1.10, SD=.04). Again, the results for Interests &
Tastes and Personality remain statistically significant with Bonferroni correction
for multiple tests.
To rule out the possibility that these differences are due to a host’s level of
experience (e.g. hosts may modify their profiles to write about specific topics
more as they host more guests), we conducted a similar analysis comparing
average hosts with superhosts, a qualification type assigned by Airbnb [Airbnb,
2019] for hosts that meet several criteria, including frequent hosting, high re-
sponse rate, and high review scores. We omit the details of this analysis for
brevity, but note that despite the fact that superhosts wrote significantly longer
profiles (a mean of 72.13 words compared to 57.74 words for non-superhosts,
t(220)=3.01, p < .001), there was no significant difference between the groups
95
in the likelihood of mentioning Interests & Tastes or Personality. This analysis
suggests that the difference between on-site and remote hosts is not due to any
difference in experience of hosting, but rather due to the expected differences
in the type of interaction that is going to take place. Taken together, the results
support uncertainty reduction theory’s central contention that people seek to
reduce uncertainty in the face of new relationships.
4.3.2 Study 2 — Self-Disclosure and Perceived Trustworthiness
Study 1 revealed that we can classify host disclosures into eight topics, and
that these topics can be reliably assessed by independent coders. An important
next question is whether the hosts were disclosing information that enhanced
perceptions of their trustworthiness. That is, do the topics that hosts disclosed the
most in Study 1 lead to higher levels of perceived trustworthiness? To examine
this question, we asked online participants to rate how trustworthy they found
each profile.
One way to operationalize the concept of trustworthiness is by using three
key dimensions: ability, benevolence and integrity [Mayer et al., 1995]. These
three dimensions are closely related but may have different effects on trust
depending on context [Colquitt et al., 2011]. Further, these dimensions are all
likely to be relevant for Airbnb profiles. In the context of Airbnb, ability refers to
domain-specific skills or competencies that the host has. Benevolence refers to
the extent to which the host is believed to want to do good to the guest beyond
profit-driven motives. Finally, integrity refers to the host adhering to a set of
moral principles and rules.
96
How might the disclosures in Airbnb profiles influence these dimensions
of trustworthiness? The Profile as Promise conceptualization of the profile
as a psychological contract implies that information provided in the profile
is an obligation by the host to a guest, namely that the information disclosed
in the profile is trustworthy and will not misrepresent the host or the host’s
home [Ellison and Hancock, 2013]. This notion of the psychological contract
suggests that hosts should be sensitive to how their promises will be evaluated for
trustworthiness by potential guests. If this is the case, then hosts should produce
promises that signal trustworthiness. We can draw on the same theoretical
perspectives we used to characterize the production of disclosures to specify
predictions about how the profile disclosures on Airbnb affect evaluations of
trustworthiness. First, uncertainty reduction theory [Berger and Calabrese, 1975]
predicts that the more information hosts disclose, the more the profile will reduce
uncertainty for profile viewers, which should enhance how trustworthy they
will be perceived. Note that more diverse information should lead to more
uncertainty reduction. That is, profiles that disclose more kinds of information
will be perceived as more trustworthy than profiles than simply say a lot about
fewer things. We therefore predict that:
H2.1 Longer and more diverse self-disclosures are perceived as more trustworthy.
Secondly, Study 1 demonstrates that hosts communicate a variety of topics in
their profiles. Signaling theory predicts that hosts use these disclosures to signal
underlying qualities or attributes that should enhance the perceptions of their
trustworthiness as a host. If the hosts have optimized their signaling behavior
for trustworthiness, then the categories they disclose most often should be the
categories of disclosure that are perceived as most trustworthy. Thus, profiles
97
A1. This person is capable of paying his/her own rent or mortgage.
A2. This person maintains a clean, safe, and comfortable household.
B1. This person will be concerned about satisfying my needs during the stay.
B2. This person will go out of his/her way to help me in case of an emergency
during my stay.
I1. This person will stick to his/her word, and be there when I arrive instead of
standing me up.
I2. This person will not intentionally harm, overcharge, or scam me.
Table 4.2: Six-item perceived trustworthiness scale.
with disclosures that were observed frequently in Study 1, including Origin or
Residence, Work or Education, Interests & Tastes, and Hospitality should be perceived
as more trustworthy than profiles that do not contain these topics.
H2.2 Self-disclosure topics used most frequently by hosts will be associated with
increased perceived trustworthiness compared to less frequent topics.
Methods
As mentioned above, we are interested in the perceived trustworthiness of host
profiles. To measure trustworthiness, we developed a six-item perceived trust-
worthiness scale on three dimensions: ability, benevolence, and integrity [Mayer
et al., 1995]. Based on items in the scale developed by Mayer et al. for an organi-
zation context [Mayer and Davis, 1999], we developed new items that measure
trustworthiness in the context of hosting. These items are shown in Table 4.2.
Items A1–A2 measure ability; items B1–B2 measure benevolence; and items I1–I2
measure integrity. When asking for profile ratings, these items were shown in a
random order.
Procedure
98
To assess the perceived trustworthiness of host profiles, we recruited crowd
workers from AMT to rate host profiles in the Experiment Dataset using the
perceived trustworthiness scale. We split the Experiment Dataset profiles into
batches of 20, and had five different workers annotate each batch. Recall that
these profiles were already labeled with the topics. We used 1,200 of the 1,234
profiles in the Experiment Dataset for this study. For each profile, workers were
instructed to rate their level of confidence regarding each of the statements, on a
scale from 0 to 100, with steps of 10. The task required that each worker only rate
one batch of the profiles to prevent any single worker’s perception from being
over-represented in the results. Workers were paid $1.00 for each task.
At the beginning of each task, we used a paraphrase question borrowed
from [Munro et al., 2010, Danescu-Niculescu-Mizil et al., 2013a] to check the
linguistic attentiveness of each worker. We re-issued the task if we received an
incorrect response to this question. To create the perceived trustworthiness score,
we calculated the perceived trustworthiness as the mean of responses for all six
items by the five workers that rated the same profile. For some analyses, we also
used three trustworthiness dimensions separately, with each score calculated as
the average of the two relevant items.
Results
We investigated the effects of profile length and diversity (H2.1), and topic (H2.2)
on perceived trustworthiness. Generally, the mean ability score of the 1,200 pro-
files was 68.82, SD=13.84; the mean benevolence score was 63.94, SD=13.97; the
mean integrity score was 66.79, SD=13.37. Note that perceived trustworthiness
scores across the three dimensions were highly correlated [pairwise Pearson’s R
99
correlation: A–B (initials): 0.86; A–I: 0.88; B–I: 0.92; p < .001].
Length, Diversity and Perceived Trustworthiness
To examine the effect of length (word count) on perceived trustworthiness,
we plot the relationship between length (x-axis, log scale) and perceived trust-
worthiness (y-axis) on each of the three trust dimensions in Figure 4.4.
Supporting H2.1, Figure 4.4 shows a clear relationship between increased pro-
file length and perceived trustworthiness scores. This relationship is confirmed
by linear regression with log transformation for profile length [b = 7.89, adjusted
R2=.38, F(1, 1198)=721.4, p < .001]. This means that when a profile doubles in
length, the perceived trustworthiness score increases by approximately 5.47,
suggesting a pattern of diminishing returns when hosts write longer profiles.
To illustrate this pattern, we divide the profiles into deciles and calculate the
average perceived trustworthiness score for each decile. Comparing profiles in
the second decile (mean word count: 13) to those in the first (mean word count:
6), mean trustworthiness score increased 18.9%; whereas comparing profiles in
the ninth decile (mean word count: 106) to those in the tenth (mean word count:
188), mean trustworthiness score only increased by 2.5%.
H2.1 also predicts that, in addition to overall length, the number of topics will
also have a positive impact on trustworthiness scores. We performed multiple
linear regression analysis with the number of topics, length as control, and
the interaction length × number of topics [adjusted R2 = 0.39, F(3,1196) = 256.6,
p < .001]. The analysis showed that the number of topics contributes to perceived
trustworthiness [b = 4.47, t(1198) = 5.54, p < .001] even when controlling for
length [log scale, b= 9.53, t(1198)= 15.03, p< .001]. There was also an interaction
100
Ability Benevolence Integrity
100
75
50
25
1 10 100 1 10 100 1 10 100
Word Count
Figure 4.4: Perceived trustworthiness increases with profile length (x-axis
on log scale).
effect between length and topic count [b = −0.95, t(1198) = −5.04, p < .001],
indicating that for shorter profiles, the number of topics increased perceived
trustworthiness even more.
Figure 4.5 visualizes the relationship between perceived trustworthiness and
the number of topics mentioned in the profile. Each line represents the density
distribution of perceived trustworthiness scores for profiles that mention a fixed
number of topics. For example, the darkest lines represent the distribution of
perceived ability, benevolence, and integrity of one-topic profiles. The figure
shows that there is variation in trustworthiness score within each topic count
bin, but as topic count increases, the trustworthiness scores also increase, and
the variations become smaller. Note that here we are not showing the effect of
profile length, which was illustrated in Figure 4.4.
Topic and Perceived Trustworthiness
101
Perceived Trustworthiness
We now analyze the effect of topic choice on trustworthiness scores. Recall
that H2.2 predicted that the topics disclosed most frequently by hosts in Study 1,
namely, Origin or Residence, Work or Education, Interests & Tastes, and Hospitality,
would also be evaluated as most trustworthy.
In our dataset, there were eight profiles that did not mention any topics,
117 one-topic profiles, 231 two-topic profiles, 239 three-topic profiles, 269 four-
topic profiles, and 336 profiles that mentioned five or more topics. We focus on
profiles that are limited to one-topic, two-topic, and three-topic combinations.
For example, looking at two-topic combinations, there are 8×7/2 = 28 different
options, although, as we show below, there are some topic combinations that
are more common than others. These 1-3 topic combinations have the most
variation, but are also simpler to study, as understanding the impact of one
single topic amid all combinations of different sizes is highly unlikely even with
1,200 profiles. We call these different combinations of topics “strategies”, and
compare the relative success of different strategies controlling for the number of
topics.
We have shown that as the number of topics increase, the trustworthiness
scores also increase. We computed one-way ANOVAS comparing the relative
effectiveness of strategies within each of the one-topic, two-topic, and three-topic
profile groups. For one-topic profiles, there was a significant effect of strategy
on ability [F(7,109) = 7.79, p < .001], benevolence [F(7,109) = 8.55, p < .001],
as well as integrity [F(7,109) = 7.36, p < .001]. For two-topic profiles, there
was no significant effect of strategy on ability [F(5,225) = 1.54, p = .18], but a
significant effect on benevolence [F(5,225) = 3.95, p < .01], as well as integrity
[F(5,225) = 2.74, p < .05]. For three-topic profiles, there was a significant effect
102
Ability Benevolence Integrity
0.05
0.04
0.03
0.02
0.01
0.00
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100
Perceived Trustworthiness
Topic Count 1 2 3 4 5+
Figure 4.5: Perceived trustworthiness score distributions for profiles with
different number of topics.
of strategy on ability [F(8,230) = 2.83, p < .01], benevolence [F(8,230) = 4.21,
p < .001], as well as integrity [F(8,230) = 2.95, p < .01].
Figure 4.6 shows the raw data for this analysis, organized by the number of
topics (panels), and three dimensions of trustworthiness scores (columns). Every
row is marked on the left with the initials of the topics in the self-disclosure
strategies (e.g. in the first row of the second panel, OW stands for the topic
combination of Origin or Residence, and Work or Education). On the right,
we show the number of profiles using this strategy (54 for OW, the most of all
two-topic strategies). The vertical lines in each row represent profiles, positioned
at the value of the perceived trustworthiness score on each dimension. The color
indicates whether the profile falls within the bottom (red), or top (green) quartile
of the profile group that used the same amount of topics (the dotted lines indicate
the bottom and top quartile boundaries).
103
Density
Ability Benevolence Integrity N
W 33
O 22
H 22
P 11
I 11
T 8
L 6
R 4
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100
Onnee-−totpoipci cS tSratrtaegtegiesAbility Benevolenceies Integrity
OW 54
HO 43
IW 21
IO 16
HP 11
Other 86
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100
Ability Two−-Btotoepnpicei cvS oStrleatrntaectgeeiegsies Integrity
IOW 32
HOW 24
IOT 20
HIO 17
OTW 16
IWT 15
HOT 14
OPW 8
Other 93
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100
ThTrheree−-ttopiic  SStrtarateteggieises
Bottom Quartile Middle Top Quartile
Figure 4.6: Comparison for different strategies, organized by the number
of topics mentioned in the profile. The dotted lines indicate the
bottom and top quartiles in each topic-count group.
104
1
There are several takeaways from Figure 4.6. Consider first the one-topic
strategies: clearly, Work or Education, Origin or Residence) and Hospitality
are the most popular, representing 66% of the one-topic profiles. Visually, it is
clear from Figure 4.6 that the most successful single-topic strategy is H, where
the profiles trending more to the right and top-quartile profiles appearing more
frequently. This visual examination is confirmed with post-hoc comparisons
using Tukey HSD tests. Among one-topic profiles, Hospitality was the best-
performing strategy, significantly trumping L, O, P, T for ability; L, O, T, W
for benevolence and integrity (p < .05, same for all the post-hoc comparisons
reported henceforth).
The second-best one-topic strategy was Interests & Tastes, outperforming L
and O for ability, and O and T for benevolence and integrity. Hosts were not
very successful writing about Life Motto & Values, although Airbnb explicitly
prompted them to, which underperforms H, I, R, W for ability; H for benevolence,
and H and R for integrity. Finally, as reflected in Figure 4.6, W outperforms O for
ability; O outperforms R for benevolence, R outperforms O and T for integrity.
None of the other pairwise comparisons were significant.
Moving on to the two-topic strategies, the dominant strategies are OW and
HO, both combinations of the most popular single-topic strategies W, O and
H, covering 42% of two-topic profiles. Interestingly, the WH strategy was not
often used (for three-topic combinations, HOW is again popular). The next two
popular strategies are IW and IO, indicating that hosts add on Interests & Tastes
as additional information. In terms of success for the two-topic strategies, post-
hoc comparisons did not indicate significant differences among strategies for
ability or integrity. However, for benevolence, HO outperforms OW and Other
105
(all other two-topic combinations that are not explicitly listed in Figure 4.6).
Finally, we see in three-topic combinations that the most common strategies
are IOW and HOW.
In terms of relative success, post-hoc comparison indicated that HOW is
clearly most successful, outperforming IWT for ability, IOT, IOW, IWT, OPW,
Other for benevolence, and IWT for integrity. In addition, HIO outperforms IOT
and IWT for benevolence. This may again be due to the high effectiveness of
Hospitality as part of the disclosure strategy, when the host is making a direct
promise to take care of the guests.
Overall, the pattern of results supports H2.2 and the prediction that profiles
with topics most frequently disclosed by hosts are also those that are evaluated as
most trustworthy. While hosts employed different combinations of topics as part
of their self-disclosure strategies, it is clear that strategies that include the most
frequently disclosed topics from Study 1 were the most successful in generating
perceived trustworthiness: Work or Education, Origin or Residence, Hospitality, and
Interests & Tastes. We now proceed to show that these trustworthiness scores are
meaningful because they have a direct impact on host choice by potential guests.
4.3.3 Study 3 — From Perception to Choice
In this section, we examine how perception of trustworthiness leads to differ-
ences in host choice. As mentioned earlier, a number of factors may influence
a potential guest’s decision to stay with a host, such as availability, price, and
characteristics of the property (e.g. location). Our primary question is whether
106
the trustworthiness signaled by profile disclosures can influence a potential
guest’s decision-making outcome, all other things being equal. To address this
question, we isolate disclosures in the profile by conducting an online experiment
to examine the effect of perceived trustworthiness on host choice. In particular,
we vary the level of perceived trustworthiness, and test the extent to which the
perceived trustworthiness of profiles influences a potential guest’s choice.
Understanding choice has important real world implications. In the face of a
potential social exchange opportunity with multiple exchange candidates, those
who portray themselves as untrustworthy can potentially be “punished”. As
shown above, the content of an Airbnb host profile affects perceived trustworthi-
ness. We know that trustworthiness differences can affect choice [Ert et al., 2016]
in other settings, and hypothesize that:
H3.1 Higher perceived trustworthiness scores for text-based host profiles predict the
likelihood of guest choice.
Methods
To test whether perceived trustworthiness affects a potential guest’s decision-
making, we employed a pairwise experiment to elicit guest response. Since we
obtained a perceived trustworthiness score for each profile in Study 2, we paired
profiles with different scores to examine if the score predicts guest’s preference
between two hosts in a pair. If the value of the trustworthiness score perfectly
predicts choice, the observed pairwise decisions we obtain from respondents
should follow the Bradley-Terry model [Bradley and Terry, 1952], which predicts
the outcome of a comparison given associated values with each participant in
107
the match.
To this end, we generated profile pairs that were comparable in length, but
with one high and one low perceived trustworthiness score. We controlled for
length for a number of reasons. Firstly, we showed above that the length is highly
correlated with trustworthiness. Choosing high- and low-scoring profiles from a
global sample is therefore likely to result in unbalanced short and long profile
pairs. We therefore used an adaptive matching method that takes length, then
score into account.
First, we ranked 1,200 annotated host profiles based on word count. Then,
from shortest to longest, we used a sliding window of roughly 240 profiles,
with steps of size 120. All profiles within each window form a group. For each
group, we calculated the bottom and top quartiles of mean trustworthiness score
(the mean of the ability, benevolence, and integrity). We then iterated through
every combination of two profiles, one from the bottom and one from the top
quartile in that group, filtering out profile pairs where one profile is longer
than the other by more than 20%. As the result of this process, we had 19,892
top-quartile-low-quartile (in the sliding window) profile pairs, representing 589
unique profiles.
The preference task for each profile pair was simple. First, each pair of profile
descriptions was shown to a respondent. For the first five seconds, the profiles
are shown but buttons were deactivated to encourage the respondent to read
the profiles before making a decision. When the buttons become activated, the
respondent click on one of the two profiles in response to the question,
“Which of the two hosts do you feel more comfortable staying with?” We
108
deliberately chose to ask about comfort, and not generally about host preference,
e.g. “Which of the hosts would you choose to stay with?” Pilot studies we ran
showed that, in making host preference decisions, people considered other per-
sonal and dyadic match factors, like their interest in staying in a location implied
in the profile. While such considerations are generally interesting, the “com-
fort” phrasing was used to focus on the effect of the trustworthiness construct.
A future study can address ecological validity by including other factors; we
show that the trustworthiness construct does impact pairwise preference based
on comfort, holding all else constant.
We recruited respondents from AMT for this task. Each respondent was asked
to evaluate 50 profile pairs in each task. To improve the quality of the results,
three pairs out of each batch of 50 were repeated in a random order, within each
batch. If the respondent provided an inconsistent answer for more than one pair,
we filtered their responses out of the analysis. Workers were paid $1.00 per task.
Experimental Results
We obtained choice results from 423 unique respondents. Consistency filtering,
and removing four responses with missing values, left us with 355 responses
consisting of 16,685 pairs of choices, representing the 589 unique profiles.
We used the Bradley-Terry model [Bradley and Terry, 1952] to evaluate pair-
wise choice. The Bradley-Terry probability model predicts the outcome of a
comparison given associated values with each participant in the match. If the
perceived trustworthiness score accurately predicts choice, it should be a good
fit to the theoretical Bradley-Terry model likelihood. Specifically, according to
109
1.00
0.75
●
●
●● Bradley-Terry
● ●
● ● ● ● ● ●
● ●
●
0.50 ● ● ●Observed
0.25
0.00
9 18 27 36 47 58 73 94 147
Mean Word Count
Figure 4.7: Observed likelihood that the host with high trustworthiness
score is preferred and the probability predicted by Bradley-
Terry model, by profile length. The error bars represent one
standard error.
the Bradley-Terry model, the probability of profile i being picked compared to
profile j is [Bradley and Terry, 1952]:
λ
P(i is preferred to j i) = (4.1)
λi +λ j
where λ is a positive-valued parameter associated with each individual
option. In our case, λ is the perceived trustworthiness score. Intuitively, the
larger the difference between λi and λ j, the higher the probability of i being
chosen (with a upper-bound of 1). In addition, if the difference between λi and
λ j is fixed, the probability of i being chosen decreases as the absolute values of λi
and λ j increase (with a lower-bound of 0.5).
Figure 4.7 summarizes, for different buckets of profile length, the effectiveness
of perceived trustworthiness in predicting choice compared to the prediction of
the Bradley-Terry model. The x-axis shows the mean word count of the profile
pairs in that length bucket, and the y-axis shows the likelihood of the profile
with the higher trustworthiness score being chosen. Both theoretical (grey) and
110
Preferred Likelihood
observed (black) likelihoods are plotted. The figure shows that for profiles in
shorter length groups (the first two groups on the left), perceived trustworthiness
predictions closely match the Bradley-Terry probability. For longer profiles,
however, the observed likelihood is lower than what is theoretically expected,
indicating less predictive power. Nevertheless, even for the longer profile, as
Figure 4.7 shows, the likelihood of choosing the top-quartile profile was higher
than chance (50%), as shown by Exact-Binomial tests for each length group
(p < .05) except for the eighth (the bin with mean word count of 94, p = .06).
The divergence of the longer profiles from the model may be due to at least
two factors. First, and more mundanely, we ran the task on AMT where work-
ers may not be incentivized to spend time reading longer profile descriptions,
skewing the results towards random chance. Second, the results may reflect the
fact that higher trustworthiness scores for longer profiles are not as predictive
for decisions. With longer profiles, other factors are more likely to be mentioned,
such as interests and tastes, that may generate specific dyadic attractions, and
play a more significant role in influencing choice. We discuss this possibility in
Section 4.5.
4.4 Step 2: A Computational Approach to Perceived Trustwor-
thiness of Airbnb Host Profiles
Again, this chapter is focused on both gaining a deeper understanding of how
people use language cues to establish trust in sharing economy, as well as pre-
dicting trust automatically. Step 1 has uncovered patterns on how host present
themselves on Airbnb (eight common topics; longer profiles for on-site hosts
111
compared to remote hosts). In addition, initial understanding has also been
established on what strategies people use to self-disclose and their relative effec-
tiveness. However, due to the limitation in data size in Step 1, we did not attempt
to build computational models to predict perceived trustworthiness directly. In
Step 2, I scale up the data collection and develop computational models to predict
perceived trustworthiness.
Specifically, we first enlarged the previous dataset to include 4,180 host
profiles annotated with perceived trustworthiness scores, the largest such dataset
to date. To enable the computational analysis, we developed several models
building on various language-based features. Using these features and models,
we evaluate a prediction task distinguishing profiles of low and high perceived
trustworthiness. In addition, we use Lasso regression to examine the factors that
contribute to higher and lower perceived trustworthiness. We discuss the results
in relation to previous research on deception [Toma and Hancock, 2012] and loan
defaults [Netzer et al., 2016], showing that the linguistic features contributing
to higher perceived trustworthiness may not always align with features that
were reported to be associated with other factors that may be indicative of actual
trustworthiness.
This part of the project builds on a number of studies using computational
approaches to study language and social interactions, enabled by new corpus and
techniques from natural language processing and machine learning. For example,
[Danescu-Niculescu-Mizil et al., 2013b] detects politeness computationally from
text by constructing new datasets of online requests; [Mitra and Gilbert, 2014]
found that in crowdfunding sites, language that makes direct promises such
as “project will be” and “pledgers will receive” is predictive of a project being
112
funded; and finally, [Netzer et al., 2016] found that language in loan requests
can help predict loan defaults.
4.4.1 Method and Dataset
Our main dependent variable here is the perceived trustworthiness of host profiles
in the context of online lodging marketplaces. Trustworthiness is defined as an
attribute of a trustee (the host). We measure perceived trustworthiness of hosts
based on their profile texts alone through a custom scale developed in [Ma et al.,
2017a], which asks potential guests about how confident they are that the host in
question is capable, benevolent, and with integrity [Mayer et al., 1995]
We used the same source dataset as well as the annotating procedure de-
scribed in Step 1. In other words, we used the Airbnb dataset collected by an
independent organization, Inside Airbnb4 on 12 U.S. major cities, and conducted
a weighted sample of profiles across all cities for 3,000 unique host profiles. We
divided profiles into 150 batches, each containing 20 profiles, and recruited three
annotators for each batch from Amazon Mechanical Turk (AMT), paying $1.5 for
each task. We required annotators to be based in U.S., adult, and with previous
approval rate of at least 90%. We also required that each worker to only perform
the task once (i.e. rate exactly 20 profiles). In this part of the project, we re-
duced the number of annotators per-profile from five in Step 1 to three, based on
analysis showing that three raters can reach a satisfying level of inter-annotator
agreement compared to five. Finally, in the exit survey, we collected additional
information about the annotators that was not collected by [Ma et al., 2017a],
including demographic information as well as their generalized trust attitude
4http://insideairbnb.com/
113
using the scale by [Yamagishi and Yamagishi, 1994].
We received 450 responses from the AMT workers and applied a series of
filtering process to exclude potential “spammers”, including checking the answer
to a linguistic attentiveness question, the standard deviation and mean of the
ratings of the same worker, and task completion time. We filtered out responses
that have very low and high (bottom and top 2.5%) standard deviation, mean,
and task completion time. We retained the rating of a host profile if it has at
least one rating after the filtering. In the end, we retained new annotations of
perceived trustworthiness of 2,980 host profiles.
After initial filtering, we performed z-score standardization on the scores
given by the same annotator, as we expected that each annotator’s scores are
subjective with different baselines for trust. Indeed, our data shows a significant
correlation between an annotator’s reported generalized trust attitude, and the
average trustworthiness scores the annotator assigned to the 20 profiles [β = .34,
t(380) = 7.08, p < .001], further justifying the decision to standardize the scores
per annotator. After standardization, we took the average of scores given to the
same profile by workers to be the perceived trustworthiness score of the profile.
To evaluate the reliability of annotations, we calculated the mean pairwise
Pearson correlation for all profiles among three raters, pooling all the data. The
average pairwise correlation is 0.49 (0.29 before standardization). Naturally,
trustworthiness is a subjective concept. Our data showed patterns similar to data
from previous research on politeness [Danescu-Niculescu-Mizil et al., 2013b]:
higher agreement at the extremes and lower agreement in the middle, which
motivated our evaluation setup as detailed below.
114
Since our annotation process is almost exactly the same as in Step 1, we
merged the new dataset with the one reported in previous work in order to boost
the amount of data available for training and testing. We performed z-score stan-
dardization on the previous dataset before merging with the dataset we newly
acquired. As a result of this merging process, we now have an extended Airbnb
host profile dataset containing a total of 4,180 profiles. The perceived trustworthi-
ness scores in our extended dataset have a mean of zero and standard deviation
of 0.8. We use this extended dataset in subsequent analysis. The extended dataset
is available online5 and contains all profile texts, perceived trustworthiness an-
notation, as well as the demographic information and generalized trust attitude
of annotators.
With the extended dataset, we set up two tasks: prediction and regression.
For the first task, our goal is to find the best model that predicts perceived
trustworthiness. As trustworthiness is a subjective concept, we set up the task
as a binary classification, following the example in [Danescu-Niculescu-Mizil
et al., 2013b]. We used logistic-regression classifiers and only top and bottom
quartile of the profiles in different profile lengths buckets in terms of perceived
trustworthiness score. For the second task, our focus is on understanding, which
we address using Lasso regression for feature selection, also using the top and
bottom quartile of the data in each length batch.
4.4.2 Predicting Perceived Trustworthiness
In this section, we set up the prediction task, and discuss features that we
construct from profile text as inputs for different prediction models, as well as
5https://github.com/sTechLab/AirbnbHosts-Extended
115
model performance.
Evaluation Setup
We split our data into two parts: a training and cross-validation set (80% of data)
that we use during model tuning, and a held-out set (the rest 20%) that is kept
separate and reserved for the final test.
We frame the prediction task for profiles of different lengths. We know
that length plays a significant role in predicting perceived trustworthiness [Ma
et al., 2017a], which is again confirmed with our extended dataset [β = .65,
t(4,178) = 55.61, p < .001]. To this end, after trimming the outliers, i.e. the
shortest and longest profiles (bottom 5% and top 5% in terms of word count),
we divided the rest of the profiles into five equal batches based on word count.
The batches, from shortest to longest profiles, have the following ranges of word
count: 6–19 words, 20–36, 37–58, 59–88, and 89–179.
Within each batch, we calculate the bottom and top quartile of the perceived
trustworthiness score. We then use logistic-regression classifiers to predict, for
profiles in these two quartiles, whether they will be in the bottom quartile (zero),
or top quartile (one), therefore only using 50% of the data in the cross-validation
set. We measure the quality of our prediction using the accuracy of the classifiers
(we are not using F1 and AUC scores as the labels are balanced).
116
Model Features
For our prediction models, we used different combinations of the features de-
scribed below. We first performed the following data pre-processing. After
removing punctuation and numbers using regular expression matching, we
converted the remaining letter words into lowercases, and removed stop words
using a union of lists of English stop words from NLTK and scikit-learn feature
extraction module, consisting of 352 stop words. Finally, we lemmatized verbs
and nouns using NLTK WordNet lemmatizer.
LIWC Features. We extracted 73 features from raw profile text (before
pre-processing) using LIWC (Linguistic Inquiry and Word Count). LIWC is
a dictionary-based text analysis tool that counts the percentage of words that
reflect linguistic process, psychological process, and personal concerns. LIWC
has been shown to predict numerous psychological outcomes [Pennebaker et al.,
2001]. We used the 2007 version of LIWC and substituted readability in LIWC
with Flesch-Kincaid grade level (extracted using the Python package textstat).
Bag-of-Words
We vectorized each of the pre-processed profiles using CountVectorizer from
the scikit-learn library. We used one-, bi-, and tri-grams and required the grams
to have appeared at least 20 times. This process resulted in 1,012 word features,
which we use in our baseline model.
Sentence Categories
In Step 1, we manually developed a set of eight sentence categories (shown
in the first column in Table 4.3) that frequently appear in Airbnb host profiles.
117
Category Accuracy F1-Score AUC
Interests & Tastes .89 .74 .92
Life Motto & Values .97 .18 .71
Work or Education .92 .79 .93
Relationships .92 .52 .88
Personality .93 .52 .89
Origin or Residence .89 .78 .93
Travel .93 .78 .95
Hospitality .86 .72 .91
Table 4.3: Performance of sentence category classification.
We created a dataset of 5,248 profile sentences tagged with categories (we used
the term “topics” in Step 1, but to avoid confusion with the term commonly used
in the context of topic modeling in the NLP community, we refer to them as
“categories” here).
Here we leverage the sentence level annotation dataset and trained eight
binary classifiers to predict whether a sentence belongs to each category. We
used the same pre-processing pipeline, and a one-gram bag-of-words model. We
set the minimum threshold of token frequency to be 10, resulting in 616 features.
We used a Bernoulli naive Bayes classifier, one for each category, and five-fold
cross validation to evaluate the performance of the classification. The accuracy,
F1-score and AUC for sentence category classification are listed in Table 4.3. The
category Life Motto & Values has the worst F1 and AUC performance due to the
extremely imbalanced label — there are very few sentences that were tagged to
belong to this category.
We applied the trained classifiers on each sentence in the extended dataset,
then adding up the classification results for sentences for the same profile into
a vector of length eight representing how many sentences in the profile were
tagged as belonging to one of the eight categories.
118
Features 6–19 20–36 37–58 59–88 89-179
words words words words words
WC 57.5% 53.8% 46.9% 43.9% 50.0%
BOW 60.8% 58.5% 62.6% 59.4% 58.8%
BOW + WC 59.1% 57.8% 63.0% 59.1% 60.2%
LIWC + WC 69.1% 58.1% 58.4% 61.8% 57.8%
Category + WC 64.0% 57.1% 62.3% 65.2% 65.3%
Category + LIWC + WC 65.9% 59.4% 65.9% 65.9% 68.0%
Best Model on Held-Out 72.4% 67.1% 51.2% 58.2% 62.5%
Table 4.4: Model performance (accuracy) summary by different length
batch. Random baseline accuracy is 50%. Models compare pro-
files of similar lengths to predict relative perceived trustworthi-
ness based on word count.
Models and Evaluation
We combine previously extracted features into different models and evaluate
their classification performance using cross-validation. We chose the simple Word
Count (WC) and BOW as baseline models. For LIWC and sentence category, we
compared the performance of models using each set of features alone, and each
plus Word Count, and found that Word Count improves performance; we only
include the WC-enhanced models here (WC did not improve on BOW model,
but BOW+WC is included here for completeness). We report the performance of
all models in Table 4.4.
The bold numbers in Table 4.4 indicate the best performing model for each
length batch. For the shortest profiles, the LIWC+WC combination achieves
the best prediction result, while longer profiles benefit from including sentence
category as features.
Evaluation on Held-Out Set
After picking the best performing model for each length batch, we re-fitted
119
the models on the entire cross-validation dataset to predict data from our com-
pletely disjoint held-out set. For the held-out set, we separated the profiles using
the word count thresholds as defined in the training stage in to each batch, and
obtained the perceived trustworthiness quartile tags using the thresholds ob-
tained from training stage. We report the accuracy of prediction on held-out set
in the last line of Table 4.4. Overall, the performance levels on the held-out set
are comparable to other text-classification work [Danescu-Niculescu-Mizil et al.,
2013b, Tan et al., 2014].
4.4.3 Factors Contributing to Perceived Trustworthiness
We conduct Lasso regression for profiles of different lengths to uncover factors
that contribute to higher and lower perceived trustworthiness. We again use the
top and bottom quartile of the data for each length batch in the cross-validation
dataset for this analysis, using the R implementation of Lasso logistic regression
(cv.glmnet) with default 10-fold cross-validation to choose the best parameter
(λ ) and using area under curve (AUC) as measure for goodness of fit. We report
features that appear in more than 10% of the profiles as well as selected to be
non-zero by Lasso in Table 4.5 and discuss the findings below.
4.5 Discussion
Sharing economy represents the third wave of the digitalization of social ex-
change. In this chapter, we took a look at how language cues establish trust in
social exchange that sharing economy platforms are designed to unlock. The
120
work presented in this chapter makes several contributions to extend our under-
standing of trust in this new context. To begin with, a new coding scheme was
introduced to categorize the language of self-disclose in Airbnb host profiles and
to annotate the perceived trustworthiness of these profiles. We find that hosts
use a variety of disclosure strategies, with some more successful than others,
suggesting that platforms can better support users to convey trustworthiness by
guiding what they disclose in profiles. In addition, computational framework
has also been developed to distinguish between Airbnb host profiles of low and
high perceived trustworthiness. We uncover features that are most predictive
to higher perceived trustworthiness, such as Hospitality sentence category, and
LIWC features social and work.
Specifically, in Step 1, the three studies reported make several contributions.
First, we developed and validated a coding scheme for self-disclosure on the
free-text portion of host profiles on Airbnb. The coding scheme describes eight
topics that covers more than 90% of current discourse in host profiles. To the
best of our knowledge, this is the first systematic coding scheme for analyzing
self-disclosure in Airbnb profiles, or more generally, for profiles related to peer-
to-peer sharing platforms.
The results of applying this coding scheme to the Airbnb profile dataset
revealed that hosts most frequently write about Origin or Residence, Work & Study
and Interests & Tastes. The least commonly disclosed topics were Life Motto &
Values, Relationships and Personality. Study 1 in Step 1 also revealed that host
type influenced the kinds of disclosures produced in host profiles, with on-site
hosts revealing more information about their Interests & Tastes and Personality
than remote hosts. These data are consistent with predictions from uncertainty
121
reduction theory [Berger and Calabrese, 1975], which predicted that on-site hosts
will disclose more information about their Interests & Tastes and Personality to
reduce potential guest uncertainty about whether they would enjoy interacting
with an on-site host.
Our studies also drew on signaling theory [Spence, 2002] to understand what
topics hosts disclose, and how guests perceive those disclosures. To assess the
implications for signaling theory, it is informative to consider the results across
both Study 1 in Step 1, which focused on the production of disclosures in host
profiles, and Study 2 in Step 1, which examined how those disclosures affected
the perceptions of trustworthiness. Signaling theory predicts that hosts will
signal their trustworthiness by disclosing more assessment signals (e.g. Origin
or Residence, Work or Study), which are more difficult to fake than conventional
signals (e.g. Life Motto & Values, Personality) [Donath, 2007]. Signaling theory
also predicts that, if hosts are optimizing their disclosures for trustworthiness,
then guests should evaluate profiles with the most frequently observed topics
as most trustworthy. The data from Studies 1 and 2 in Step 1 largely confirmed
both of these hypotheses: hosts disclosed more assessment signals than conven-
tional ones, and guest perceived profiles with more assessment signals as more
trustworthy.
There were, however, some important exceptions to the theoretical predic-
tions. Certain strategies, such as demonstrating Hospitality or sharing one’s
Interests & Tastes, proved to be more successful than expected for conventional
signals while other strategies, such as providing only one’s Origin & Residence,
proved less successful. The language of Hospitality is more successful potentially
because it provides more information about expected interactions, therefore re-
122
ducing uncertainty in future exchange and increasing trust. Providing a welcome
or greeting, or providing reasons for hosting, alone but preferably in combination
with more assessment signal disclosures (e.g., Hospitality combined with Origin
& Residence), was an important strategy that had a strong and positive effect
on perceptions of trustworthiness. These findings suggest that signaling with
conventional signals but that provide information about one’s hospitality or
interests can enhance trustworthiness in Airbnb profiles.
Finally, we demonstrated that perceived trustworthiness matters for decision-
making in this context. The perception of trustworthiness from Study 2 in Step 1
predicted participants’ decisions in a forced choice experimental task in Study 3
in Step 1, especially for profiles that are relatively short (less than 20 words).
We also showed that when profiles are short, perceived trustworthiness almost
perfectly predicts choice, whereas when the profile length increases, other factors
appear to also influence choice. This may suggest a nuanced role of trust in
decision-making — there is a threshold of trust that is needed to pass muster,
but other factors (e.g. homophily [McPherson et al., 2001]) may weigh in once
trustworthiness is no longer the issue.
This research suggests that the Profile as Promise framework [Ellison et al.,
2012] is a useful approach for understanding how hosts and guests produce
and evaluate disclosures in Airbnb profiles. Hosts disclosed information about
themselves that they perceived as relevant and of interest to potential guests, and
their promises were evaluated based on their trustworthiness, as predicted by
signaling theory and uncertainty reduction theory. This study suggests that the
concept of a promise, or psychological contract, can be usefully applied beyond
online dating profiles [Ellison and Hancock, 2013] and résumés [Guillory and
123
Hancock, 2012] to peer-to-peer sharing platforms such as Airbnb.
At the same time, in Step 2, the computational approaches allow us to gain
more detailed insights on what factors and language features contribute to higher
perceived trustworthiness. The key factors contributing to higher perceived trust-
worthiness are Hospitality sentence category, and LIWC features social and work.
The sentence category Hospitality contributes to higher perceived trustworthi-
ness for profiles longer than 37 words. The effectiveness of hospitable language
strengthens the findings of [Ma et al., 2017a]. The social LIWC category also con-
tributes to higher perceived trustworthiness, potentially through the mechanism
of uncertainty reduction [Berger and Calabrese, 1975]. Providing information
about one’s social relationships can make hosts appear more “real”. Finally,
LIWC feature work predicts higher perceived trustworthiness for all profiles
shorter than 59 words, potentially also through the mechanism of uncertainty
reduction.
In contrast, LIWC feature leisure and sentence category Interest & Tastes are
contributing negatively to perceived trustworthiness for profiles between 6–19
words and 37–58 words respectively. The negative effect of these features may
suggest a separation between the need of sociability and the ability to provide
standard goods and services in sharing economy.
Comparing these features that were found to be significant in our work with
previous research, we uncover a potential discrepancy between the language that
is perceived to be trustworthy, and the actual trustworthiness of individuals. In
terms of perception, as we see in our work, and previous work on crowd funding,
LIWC features social and work contribute to higher perceived trustworthiness or
higher likelihood of a project being funded [Mitra and Gilbert, 2014]. However,
124
in terms of actual trustworthiness, social language is found to be associated with
higher loan default rates [Netzer et al., 2016]; and in online dating profiles, online
daters used more work related words [Toma and Hancock, 2012] when their
photos are less accurate. Expanding on the discrepancy between perceived and
actual trustworthiness would be important future work.
4.5.1 Design Implications
Our findings have direct implications for improving the design of profile pages
on sharing economy platforms, with the view of encouraging trustworthiness
and improving the rate of transactions. Our results suggest that hosts should be
encouraged to disclose more information, and that this information should come
from a diverse set of the eight categories identified in the coding scheme from
our study. With knowledge of the profile features that may promote trust, inter-
faces for creating and editing profile text could encourage individuals to write
more, and focus on the key categories exposed above. Automatic text analysis
mechanisms could also be used to classify text into categories, and suggest other
topics to improve breadth and ultimately perceived trustworthiness.
4.5.2 Limitations
There are some important limitations to this work. First and foremost, we opted
to prioritize our theoretical understanding of trustworthiness in profiles, over
developing an ecologically valid measure of the profile text’s effect on host choice.
As mentioned above, host choice on Airbnb can be impacted by many factors,
125
including (most trivially) the price and characteristics of the rental property.
Nevertheless, the experiment we ran on host choice allowed us to make causal
claims regarding the profile text’s impact on guest decision-making.
A related limitation of the work is the fact that it ignores dyadic and dynamic
determinants of trustworthiness. A key mechanism of uncertainty reduction
theory involves dyadic reciprocity and exchange [Berger and Calabrese, 1975].
In this work, we only examined a single-sided, one-time disclosure by hosts. It
would be important to consider the effect of the dyadic properties of hosts and
guest, and how they relate to trustworthiness and trust. Understanding how
impressions of perceived trustworthiness form and evolve through conversations
between hosts and guests would be another complex and interesting problem to
tackle.
Our dataset only includes U.S. large cities. As a result, the findings may
not generalize to hosts in smaller cities, though nothing in our findings would
necessarily suggest that this would be the case. We did not consider gender
and cultural differences in this work, either. In a preliminary investigation, we
inferred the gender of hosts from their first names, but did not find significant
differences in disclosures between hosts of different gender. Future work can
dive deeper into patterns of self-disclosure by individuals of different gender
and cultures, potentially helping to combat discrimination or potential biases
known to exist on sharing economy platforms [Thebault-Spieker et al., 2015].
While this work uses quantitative approach, another approach would have
used other qualitative methods, such as interviews with hosts about their profile
construction strategies. For example, how do Airbnb hosts present a trustworthy
facade while balancing other important aspects (e.g. privacy)? Other research
126
has qualitatively examined the experiences of hosts [Lampinen and Cheshire,
2016] but to date has not considered profile construction work.
A key question that requires future work is whether our findings are unique
to the context of lodging in sharing economy, more specifically to Airbnb, or
maybe they apply more generally in sharing economy platforms. While we
believe some features we identified, for example some LIWC features in the
different models, apply more generally, other features are context specific. For
example, the sentence category Hospitality is specific to the lodging context,
though at the same time a version of it can transfer to other domains (e.g.,
promise of service). Future work can expand our results to other instances of
sharing economy platforms.
Finally, because we relied on crowd workers to provide ratings of perceived
trustworthiness, our annotation and the resulted algorithms trained on the
annotated data run the risk of having potential bias. For example, it is possible
that people have stereotypes towards certain professions, such as artists. When
that is revealed in self-disclosure in profiles, it is possible that such stereotypes
will be penalized in terms of perceived trust. Another potential bias is that people
might perceive women hosts as more trustworthy. Future work can look into
whether systematic bias exist in the annotation.
4.6 Conclusion and Extensions
In this chapter, we conducted a two-step inquiry of how language establishes
trust through self-disclosure in the context of online sharing economy platform —
Airbnb. Through the computational modeling of Airbnb host profiles, we show
127
our frameworks can distinguish between Airbnb host profiles of low and high
perceived trustworthiness. In addition, we uncover language cues that are most
predictive to higher perceived trustworthiness, including cues of hospitality, cues
related to social context, and cues related to work. This shows that in addition to
uncertainty reduction, signaling plays a very important role in self-disclosure to
establish trust through language on sharing economy platforms.
Taken together, the last and this chapter represent the first focus of networked
trust: cues in Computer-Mediated Communication, including image cues and
language cues. As shown in both chapters, we can develop custom algorithms
to predict high or low quality images, or to predict profiles that are perceived
as high or low in trustworthiness. These algorithms can be fitted to provide
feedback for users to take better images, write better profiles, or even generate
profiles altogether, to optimize for trust.
Such potential for algorithmic augmentation of online presentation raises
new questions about the reliability of cues in Computer-Mediated Communica-
tion. As a result of algorithmic mediation, what used to be Computer-Mediated
Communication (CMC) is turning into AI-Mediated Communication (AI-MC):
interpersonal communication not simply transmitted by technology but aug-
mented — or even generated — by algorithms to achieve specific communicative
or relational outcomes. In AI-Mediated Communication, an AI system oper-
ates on behalf of the communicating person, e.g., by augmenting, generating
or suggesting content. AI-Mediated Communication is distinct from traditional
CMC technologies that primarily transmit messages, and from typical machine-
authored texts that do not represent a person.
In one of the follow-up works to the study presented in this chapter, we
128
observed the effects of AI-Mediated Communication in the first attempt to con-
ceptualize AI-Mediated Communication [Jakesch et al., 2019]. Through a series
of three online experiments, we examine how the belief that a computer system
has generated a host’s profile changes whether the host is seen as trustworthy
by others. We observed that (1) when people are presented with all AI-generated
profiles they trust them just as they would trust all human-written profiles; (2)
when people are presented with a mixed set of AI- and human-written profiles,
they mistrust hosts whose profiles they believe were generated by AI.
As algorithms continue to gain in importance in mediating our communi-
cation online, it is important to understand the theoretical and practical im-
plications of algorithmic mediation, especially on interpersonal trust. Future
studies can continue to investigate AI-Mediated Communication’s impact on
interpersonal trust as new applications continue to appear on the market.
129
130
Profile Length Positive Features Negative FeaturesSentence Category LIWC Sentence Category LIWC
6–19 words (Not included) readability, comma, article, con- (Not included) word per sentence, exclamation
junction, social, affect, causation, mark, present tense verb, adverb,
health, work, achieve quantifier, human, tentative, per-
ceptual process, sexual, relativity,
motion, space, leisure
20–36 words Travel we, social, positive emotion, — parenthesis, adverb, prepositions,
cognitive process, sexual, work, perceptual process
home
37–58 words Work or Ed- parenthesis, we, article, auxiliary Interests & comma, dash, exclamation mark,
ucation, Re- verb, past tense verb, social, fam- Tastes, Personal- period, negation, quantifier, in-
lationships, ily, friend, certain, work ity sight, motion, leisure
Hospitality
59–88 words Relationships, we, social — —
Hospitality
89–179 words Hospitality social, inclusive — —
Table 4.5: Factors contributing to higher and lower perceived trustwor-
thiness by different length batch. LIWC categories “we” and
“social” are important positive indicators, potentially through
social warranting. The LIWC “sexual” category contained mostly
the word “love”.
CHAPTER 5
NETWORKS OF TRUST: WHEN DO PEOPLE TRUST THEIR SOCIAL
GROUPS?
5.1 Introduction
Last two chapters covered the first focus of networked trust — cues in Computer-
Mediated Communication. This chapter presents a case study exploring the
second focus of networked trust: trust in social exchange that are embedded in
social networks. In particular, I study how social networks contribute to trust in
the context of Facebook groups.1 Again here we use a survey-based measure to
assess the perceived trustworthiness of other members in a social group.
As one of the most prominent social network sites (SNSs), Facebook is among
the second wave of the digitalization of exchange — the digitization of social
relationships. SNSs enable users to articulate and make visible their social
networks [Boyd and Ellison, 2007], and they require that people trust each
other with their personal information. SNSs brought new challenges about
trust, especially in relation to privacy [Marwick and Boyd, 2011], but they also
present new opportunities to understand how social structure contributes to
trust. For example, people’s friendship network structure has been shown to
be informative for predicting romantic relationships [Backstrom and Kleinberg,
2014].
This chapter leverages the opportunities to study trust in social groups on
SNSs. Social groups are important social structures through which communities
1This work was published at CHI 2019 as When Do People Trust Their Social Groups? [Ma et al.,
2019a]
131
are formed. On Facebook, billions of people engage with social groups every
month. Trust is attributed to contribute to the success of social groups by encour-
aging people to interpret others’ actions and intentions favorably, thereby facilitat-
ing cooperation and a sense of community [Gambetta, 1988, Misztal, 2013, Uzzi,
1996, Bachmann, 2001, Dirks, 1999, Preece and Maloney-Krichmar, 2005]. In
groups, trust increases member satisfaction and task performance [Walther and
Bunz, 2005], reduces conflict [Gambetta, 1988, Walther and Bunz, 2005], and pro-
motes effective response to crisis [Meyerson et al., 1996]. Previous research has
examined how different factors such as size [Brewer, 1991, Denters, 2002, Zelmer,
2003], group cohesiveness [Hogg, 1993], and activity [Walther and Bunz, 2005]
may impact people’s trust in their social groups, both online [Holtz et al., 2017]
and offline [Rotter, 1971]. However, previous studies tend to be small in scale,
limited to specific contexts (e.g., online marketplaces), or only consider a spe-
cific type of group (e.g., organizations [Mayer et al., 1995, Colquitt et al., 2007]).
Studies that address these three limitations may enrich our understanding of
how trust is formed in social groups more generally. In particular, as people in
a group are connected in different ways, social network structure of the group
may play an important role in structurally determining how much people trust
the group.
In this work, we build on rich prior literature on trust to present a frame-
work for predicting an individual’s trust in a social group, and examine how
differences at the individual and group levels predict that trust, especially the
social network structure of groups. We focus our analysis on Facebook Groups2,
a Facebook feature that “allows people to come together to communicate about
shared interests” [Facebook, 2018]. As of May 2018, 1.4 billion people use Face-
2We use “Groups” to refer to the Facebook product, and “groups” to refer to actual social
groups on Facebook.
132
book Groups every month [Perez, 2018]. By combining a survey (N=6,383 valid
responses) of individuals using Facebook Groups with aggregated behavioral
logs, we are able to investigate, across a diverse sample, how an individual’s
trust in a group relates to characteristics of the individual, the group, and the
individual’s membership in that group.
The survey asked individuals about their general attitudes towards others and
trust towards a Facebook group that they were a member of. While prior work
has shown that an individual’s general propensity to trust others influences their
trust in a particular group [Boss, 1978,Butler Jr, 1999,Ridings et al., 2002,Ferguson
and Peterson, 2015], we additionally examine the role of other individual-level
differences (e.g, general attitudes towards risk-taking).
We combine these survey results with aggregated behavioral and descriptive
data on Facebook Groups. This allows us to study the role of five categories of
features that characterize either the group or the respondent’s relationship with
the group, based on prior literature: (1) basic properties of the group (e.g., size,
membership privacy policy) [Kraut and Fiore, 2014]; (2) group category [Denson
et al., 2006]; (3) group activity [Kraut and Fiore, 2014]; (4) group homogene-
ity [Moser et al., 2017]; and (5) the friendship-network structure of the group
[Holtz et al., 2017].
We find that these variables robustly predict participants’ trust in a particular
group, with both individual and group characteristics contributing predictive
value (adjusted R2=0.26). In particular, an individual’s trust in a group was most
strongly predicted by their general perceived social support, the group’s average
clustering coefficient, and their degree centrality in the group. We also show that
trust in a group can be estimated using only observational data.
133
While these results support previous findings showing that intragroup trust
decreases with increasing group size and increases with membership restric-
tion [Zelmer, 2003, Denters, 2002, Brewer, 1991, La Macchia et al., 2016, Moser
et al., 2017], we find that these trends only hold up to a certain point. When the
size of a group exceeds 150 members (roughly Dunbar’s number, or the expected
cognitive limit beyond which social relationships are difficult to maintain [Dun-
bar, 1992]), the membership policy of the group (public v.s. private) ceases to
play a predictive role. Moreover, in deciding how much to trust a group, we
show that group size matters less to individuals with a higher general propensity
to trust.
Further, previous work suggests that people trust groups in which they are
more active [Cartwright and Zander, 1953], but we find that only certain types of
activities are associated with trust: people “like” and “comment” more in groups
they trust but do not necessarily post more, suggesting that trust is associated
more with directed communication than with information sharing.
Finally, we show that trust in groups is associated with both individual- and
group-level outcomes. Increased trust leads to individuals being more likely to
form friendships with other members of the group, but is also associated with
the group being less likely to grow larger in size.
In summary, we (1) present results of a large survey of individuals’ trust
attitudes towards their social groups (6,383 unique groups); (2) examine how
characteristics of both the individual and group contribute to trust in a group;
and (3) show how this trust affects future individual- and group-level outcomes.
A deeper understanding of how these factors collectively contribute to trust in
groups can better equip communities to foster trust among their members.
134
5.2 Related Work
5.2.1 Determinants of Trust in Groups
What contributes to trust in groups? Here we review relevant literature that
guide the selection of our feature sets in predicting trust in groups.
Individual Differences
Trust in groups can be mediated by one’s disposition to trust others, as it cor-
relates with one’s initial intentions to trust a group, especially in ambiguous
situations [Gill et al., 2005]. A disposition to trust can positively impact trust in
different settings, including trust between individuals [Yakovleva et al., 2010],
in communities [Ridings et al., 2002], in organizations [Kantsperger and Kunz,
2010], or in online services [Wu et al., 2010]. Similarly, a disposition to trust
increases trustworthiness evaluations given to Airbnb hosts [Ma et al., 2017a],
though in other settings, a disposition to trust was not associated with trust in
peer sellers [Jones and Leonard, 2008].
Past work also suggests an inverse relationship between risk aversion and
trust [Abrahao et al., 2017] — the more comfortable an individual is with taking
risks, the higher the trust they have in groups.
Further, prior literature treats membership of voluntary associations as an
indicator of trust [Putnam, 2000,Putnam, 1993]. Thus, greater in-group loyalty, as
well as perceived social support from others, should both be linked with higher
trust in groups due to increased group participation and perceived social capital.
135
Group Characteristics
Trust in groups may also stem from basic properties of the group such as its
size [Brewer, 1991, Denters, 2002, Zelmer, 2003]. For instance, experiments have
shown that people identify more strongly with smaller groups [Simon and Brown,
1987]. In addition, groups that have existed for longer periods of time should
also be trusted more, as they have more time to develop established norms and
culture that are beneficial for group trust. Past research has also described how
secrecy can build community [Fine and Holyfield, 1996] and shown that group
cohesiveness promotes trust [Stokes, 1983]. Recent qualitative work on Facebook
Groups also suggests that by making membership exclusive and screening new
members, trust can be enhanced [Moser et al., 2017].
Homogeneity, which relates to cohesiveness, may also contribute to trust.
People tend to be closer to and trust others who are similar to them [McPherson
et al., 2001]. Research has also found a relationship between gender and age
homophily and increased trust [Ahmad et al., 2011, Abrahao et al., 2017].
Higher levels of group activity are also linked with greater trust [Cartwright
and Zander, 1953, Walther and Bunz, 2005]. Increased social interaction provides
“opportunities for people to get acquainted, to become familiar with one another,
and to build trust” [Ren et al., 2007], thus leading to higher familiarity, and in
turn, greater trust [Gulati, 1995].
Network Characteristics
Beyond group characteristics mentioned above, the overall structure of relation-
ships between individuals in the group, as well as the individual’s embeddedness
136
the group’s social network may mediate trust. A person’s number of friends
and the connections among these friends can both increase the likelihood of
them joining a community [Backstrom et al., 2006, Ugander et al., 2012]. As
dense networks promote cooperation and social norms, they are also likely to
be associated with increased trust [Coleman, 1988]. In buy-and-sell groups on
Facebook, network density and the degree centrality of the seller are positively
correlated with an intention to transact, which may signal higher trust in the
group [Holtz et al., 2017]. Our work uses similar features but directly measures
trust via a survey, and considers the role of network features within a much large
set of variables.
This rich prior literature motivates our analysis in this work, in which we con-
duct a large-scale survey and analyze behavioral data to show how individual-
and group-level differences help predict trust in groups. Our research ques-
tions are as follows: (a) Can a baseline model that accounts only for individual
attitudes predict trust in groups? (b) What is the relative contribution of the
different sets of group-level features (basic group properties, group category,
activity, homogeneity, and structural properties) on trust in groups beyond the
baseline model?
5.3 Methods
In this work, we conducted a survey of 10,000 respondents to a random sample
of active Facebook Groups users in the U.S. People were invited to participate in
the survey via an ad on Facebook. The survey was designed to measure both
individual attitudes as well as trust in one of the randomly selected Facebook
137
groups of which they were members. We augmented this survey data with
self-reported demographic data such as age and gender and server logs of these
individuals’ activity and friendships on Facebook. All log data was de-identified
and analyzed in aggregate on Facebook’s servers; researchers did not view any
identifiable data nor access any specific posts in any groups. The study was
approved by an internal Facebook board as well as Cornell’s Institutional Review
Board under protocol #1805008006.
5.3.1 Sampling
The survey was issued to unique individual-group pairs. We used the following
sampling strategy to identify eligible survey candidates. First, we identified
Facebook groups that had at least five members and that had a majority of their
members located in the U.S. We then identified people in the U.S. who belong to
at least one of these groups, and that had at least one interaction (e.g., creating,
liking, or commenting on a group post) in the past 28 days. We then sampled
eligible individual-group pairs, de-duplicating by both individual and group at
random. The sampling was also stratified by group size (the number of members
in the group) to better capture behavior across both smaller and larger groups.
We note the following bias introduced by our sampling method: compared to
all individuals who actively used groups in the past 28 days, our participants
tended to be 8.7% older and were 17.5% more likely to be women.
138
5.3.2 Survey Design
The survey consisted of two sections: a section on individual differences regard-
ing the participant’s general attitudes towards others, including disposition to
trust and related concepts; and a section on trust in a specific Facebook group.
Each section had four items, shown in Table 5.1. The order of questions was
randomized within a section. Participants were asked to report the extent to
which they agreed or disagreed with each statement on a five-point Likert scale.
For the section on general attitudes towards others, we measured disposition
to trust through an adaptation of the generalized trust question in the World
Values Survey [WVS, 2018]. The original question elicits a dichotomous response,
worded as: “Generally speaking, would you say that most people can be trusted
or that you need to be very careful in dealing with people?” We instead used
a more granular five-point Likert scale, which has been shown to be more reli-
able [Miller and Mitamura, 2003]. We also included measures of concepts related
to disposition to trust reported in previous literature, including general social
support [Barrera Jr and Ainlay, 1983, Vigoda-Gadot and Talmud, 2010, Hether
et al., 2014], in-group loyalty [Van Vugt and Hart, 2004], and risk aversion [Miller
and Mitamura, 2003].
To measure an individual’s level of trust in a Facebook group of which they
were a member, we created a four-item scale to measure trust in groups (section
two in Table 5.1), based on previous literature. This scale is based on the frame-
work of ability, integrity, and benevolence by Mayer et al. [Mayer et al., 1995] and
Schorman et al. [Schoorman et al., 2007], and also adapts measures from several
interpersonal trust scales including Rotter Interpersonal Trust Scale [Rotter, 1971],
the Specific Interpersonal Trust Scale [Johnson-George and Swap, 1982], and a
139
newer “predisposition to trust” scale [Ashleigh et al., 2012].
In addition, to better understand what people use the group for, we also
asked participants to use the taxonomy below to describe the group category:
• Friends & Family: e.g., close friends, extended family
• Education & Work: e.g., college, job, professional
• Interest-Based: e.g., hobby, book club, sports
• Identity-Based: e.g., lifestyle, health, faith, parenting
• Location-Based: e.g., neighborhood or local organization
• Other
These categories were identified in previous qualitative research, where we
surveyed people who used Facebook Groups and asked them to describe a group
they were part of (e.g., “my family”). In our work, participants were requested
to select all categories that applied to the group they were surveyed on, and we
treated each group category as a binary variable. In our sample, around 34% of
the groups were tagged as interest-based groups (most common), followed by
20% friends & family groups. The first five categories capture most of the groups
(covering 89%).
5.3.3 Data and Statistical Approaches
In addition to data from the survey, we examined properties of groups including
their sizes and membership privacy policies. For each group, we also looked at
an individual’s activity in the group (e.g., time spent, likes, comments, and posts
140
General Attitudes Towards Others
Disposition to trust Most people can be trusted.
General social support There are people in my life who give me support and
encouragement.
General risk attitude I’m willing to take risks.
General in-group loyalty I would describe myself as a “team player”.
Trust in a Group
Care Other members of the group care about my well-being.
Reliability Other members of this group can be relied upon to do what they
say they will do.
Integrity Other members of this group are honest.
Risk-taking I feel comfortable sharing my thoughts in this group.
Table 5.1: Trust in groups survey items. Participants reported the degree
to which they agreed or disagreed to each of the survey items on
a five-point Likert scale.
made), the group’s overall activity, as well as group members’ friendships with
each other.
Out of the 10,000 survey responses we received, we filtered responses based
on the completeness of the survey, as well as availability of self-reported and log
data. In the end, we retained 6,383 responses for our main analysis.
The main statistical techniques we used were multiple linear regression,
random forests, and logistic regression. We first built a baseline model predicting
trust in groups using variables capturing individual-level differences. Then,
we identified five different sets of group-level features, conducted analysis on
how much each set of feature improved the baseline model, and interpreted the
relationship between each feature and trust separately. We next combined all
features in a random forest model and compared the importance of each set of
features in the combined model. Finally, we used logistic regression to predict
group outcomes such as the densification of the friendship network within the
141
Variable Mean SD Correlations
1 2 3
General Attitudes Towards Others
1. Disposition to trust 3.33 1.07
2. General risk attitude 3.68 0.99 0.18***
3. General in-group loyalty 4.54 0.80 0.23*** 0.16***
4. General social support 4.34 0.84 0.24*** 0.18*** 0.36***
Trust in Groups
1. Care 3.90 1.08
2. Reliability 4.05 0.98 0.62***
3. Integrity 4.20 0.95 0.60*** 0.67***
4. Risk-taking 4.09 1.06 0.59*** 0.54*** 0.56***
Note: ∗p<0.05; ∗∗p<0.01; ∗∗∗p<0.001
Table 5.2: Descriptive summary of survey measures, including general
attitudes and trust in groups. Sparklines represent the histogram
of each measure. (N=6,383)
group. When appropriate, we log-transformed the data (e.g., group size) and
note the transformation when reporting coefficients.
5.4 Results
Trust in groups was measured in our survey across four dimensions: care, re-
liability, integrity, and risk taking. As shown in Table 5.2, these dimensions of
trust in groups are highly correlated (ρ ≥ 0.54; Cronbach’s α = 0.86). Thus, we
defined a composite “trust in groups” score as the mean of all four dimensions
and report findings with respect to this composite score.
142
Dependent variable:
Trust in groups composite score
(1) (2) (3) (4) (5)
(Intercept) 4.07∗∗∗ 3.51∗∗∗ 3.18∗∗∗ 2.46∗∗∗ 1.98∗∗∗
(0.04) (0.05) (0.06) (0.07) (0.08)
Age −0.001∗ −0.003∗∗∗−0.002∗∗ −0.001∗ −0.001∗
(0.001) (0.001) (0.001) (0.001) (0.001)
Female 0.09∗∗∗ 0.06∗∗ 0.08∗∗∗ 0.05∗ 0.02
(0.02) (0.02) (0.02) (0.02) (0.02)
Disposition to trust 0.19∗∗∗ 0.17∗∗∗ 0.13∗∗∗ 0.11∗∗∗
(0.01) (0.01) (0.01) (0.01)
Risk attitude 0.09∗∗∗ 0.07∗∗∗ 0.05∗∗∗
(0.01) (0.01) (0.01)
In-group loyalty 0.21∗∗∗ 0.16∗∗∗
(0.01) (0.01)
Social support 0.19∗∗∗
(0.01)
Adjusted R2 0.003 0.06 0.07 0.11 0.14
Note: ∗p<.05; ∗∗p<.01; ∗∗∗p<.001
Table 5.3: Baseline model predicting trust in groups using demographics,
disposition to trust, risk attitude, in-group loyalty, and social
support. (N=6,323 after removing missing age and gender obser-
vations)
5.4.1 Individual Differences and Trust
We start by predicting trust in groups using individual attitudes as well as
demographic information (see in Table 5.3), which prior literature has associated
with differences in one’s disposition to trust [Taylor et al., 2007].
Demographics
We found that demographic factors such as the age and gender of participants
capture almost no variance of trust in groups (see Model 1 in Table 5.3). This
result partially contrasts with the prior work that found a relationship between
143
these demographic factors and one’s disposition to trust [Taylor et al., 2007]. To
better understand this result, we tested a model that used demographic variables
to predict participants’ disposition to trust rather than trust in groups. While we
found that older people were more trusting than young people (β=0.006, p<.001)
and women were more trusting than men in general (β=0.12, p<.001), very
little variance in disposition to trust is explained by these demographic factors
[R2=0.01, F(2,7174)=36.1, p<.001]. In other words, demographic characteristics
explain neither an individual’s disposition to trust nor their trust in groups.
General Attitudes Towards Others
How do an individual’s general attitudes towards others predict their trust in
groups? Corroborating prior work, one’s general disposition to trust significantly
predicts one’s trust in groups (see Model 2 in Table 5.3). However, other factors
also play significant roles (Models 3–5 in Table 5.3). Notably, the individual’s
perceived social support (β=0.19, p<.001) and their general stated in-group
loyalty (β=0.16, p<.001) contributed more to the prediction of trust in group than
one’s disposition to trust (β=0.11, p<.001). A willingness to take risks (β=0.05,
p<.001) was least predictive. Altogether, these factors capture a significant
amount of the variance in trust in groups (adjusted R2=0.14).
5.4.2 Group Differences and Trust
To understand the relationship between group characteristics and trust in groups,
we identified five distinct sets of group-level features (see Table 5.4). In this
section, we measure the incremental predictive value of each of these sets of
144
Feature Set Features
Basic Properties (5) Group size, privacy type, group tenure, number of ad-
mins/moderators
Category (5) Self-reported group category
Activity (6) Group-level and participant-group-pair level time spent,
number of posts, number of likes or comments
Homogeneity (3) Diversity of group age, gender, and similarity between par-
ticipant and group average
Structural (5) Network density, average clustering coefficient, participant
degree centrality, cliquishness of participant’s friends in the
group, average number of mutual friends with group mem-
bers
Table 5.4: Five sets of group-level features used for predicting trust in
groups.
group-level features, after controlling for the individual differences discussed
above. Here, we use “baseline model” to refer to a model that only includes
the individual differences (Model 5 in Table 5.3). For each feature set, we add
the features as independent variables in the multiple linear regression model to
the baseline model. In each subsection, we report how much the model gains
from the additional features. We validated that the coefficients of the individual
differences features in the baseline do not change significantly when we include
each new feature set.
Basic Group Properties
The first set of group-level features consisted of group size, privacy type, group
tenure (how long a group has existed), the number of group admins, and its
number of moderators. Adding these features to the baseline model increased
the model’s adjusted R2 by 0.08 (p<.001). The most significant predictor of trust
was group size. Consistent with previous work on trust and group sizes [Brewer,
145
1991, Denters, 2002, Zelmer, 2003], people had lower trust in bigger groups
(β=−0.15 on log scale, p<.001).
Apart from a group’s size, a group’s privacy type can also affect perceptions
of trust. On Facebook, group admins can set the group to be “public”, “closed”,
or “secret”. Public groups are accessible to non-members, while closed and
secret groups are only accessible to current members; closed groups differ from
secret groups in whether their existence is known to non-members. We found no
significant differences between closed and secret groups, so we analyzed them
together as “private” groups.
Controlling for group size (public groups are 68% larger than private groups),
we found that people trusted public groups less than private groups (β=−0.07,
p<.01), as suggested in prior work [Moser et al., 2017].
Notably, we found an interaction effect between group size and privacy type
in predicting trust (β=0.04, p<.01): the larger the group, the smaller the difference
there is between trust in private and public groups. To see how quickly this
difference between group types dissipates, we conducted a series of t-tests in
which we compared the mean difference in the trust composite score between
public and private groups above a certain size threshold, starting from 10 in
increments of 10. These tests show significant differences between groups larger
than the threshold until the threshold exceeds 150, where we no longer observe a
significant difference between public and private groups (p>.05).
Incidentally, this size threshold roughly corresponds to Dunbar’s number
— the maximum number of stable social relationships a person can maintain
due to limitations in cognitive resources [Dunbar, 1992]. Smaller private groups
146
Care Reliability Integrity Risk Taking
4.5
4.0
3.5
3.0
2.5 5.0 7.5 2.5 5.0 7.5 2.5 5.0 7.5 2.5 5.0 7.5
Group Size (Log)
Disposition to Trust Low High Privacy Type Public Private
Figure 5.1: The relationship between trust in groups and group size, for
each dimension (panels), across groups of different privacy
types (line style) and individuals with different propensity to
trust (line color). Dunbar’s number (150) is marked by a vertical
red dotted line.)
provide control and exclusivity over membership, thus allowing members to
foster a shared sense of identity [Moser et al., 2017]. Once the group becomes too
big, that shared identity might be lost, resulting in no difference between large
groups that are public or private.
Figure 5.1 summarizes the impact of group size on trust in public and private
groups, as well as the effect of an individual’s disposition to trust (we consider
composite scores >3 to be high and <=3 to be low). The figure shows that having
a high disposition to trust (black lines) and a group being private (solid lines)
are both factors that contribute to trust in groups. But while the effect of privacy
decreases with size (dashed and solid lines cross), the reverse is true for an
individual’s disposition to trust. An interaction effect between group size and
individual’s disposition to trust (β=0.01, p<.01) shows that people with a greater
disposition to trust others were less sensitive to changes in group size (visually
represented by gentler slope of black lines compared to blue ones in Figure 5.1).
147
Trust in Group
Other basic group properties also relate to trust. Longer group tenure predicts
higher trust (β=0.04 on log scale, p<.001), potentially because older members
have more stable group relationships and are more familiar with other group
members [Walther and Bunz, 2005]. The number of admins also predicts higher
trust (β=0.10 on log scale, p<.001). This finding is consistent with previous work
that found that groups with more admins tended to survive longer than groups
with fewer admins [Kraut and Fiore, 2014]. The number of moderators is a much
weaker predictor of group trust.
Group Category
As previously described, participants in our survey labeled groups as belonging
to one or more of six categories. Including group category as multiple binary
variables to the baseline model significantly improved trust predictions (p<.001),
increasing the model’s adjusted R2 by 0.05. To illustrate differences in trust across
these categories, we also conducted an ANOVA and plotted the average trust in
groups by category in Figure 5.2. Groups marked as “other” were excluded from
this analysis. Post-hoc Tukey tests showed that people trust friends & family
groups the most, followed by identity-based and education & work groups
(p<.001). They trust interest- and location-based groups least (p<.001).
Why does trust differ by group category? For friends & family groups,
high trust is a strong sign of bonding social capital [Putnam, 2000]. Identity-
based groups (e.g., parenting groups) and education & work groups elicit trust
by establishing a shared identity among group members [Moser et al., 2017].
Finally, interest- and location-based groups may represent less bonding and more
bridging social capital [Granovetter, 1985], especially for information sharing.
148
4.50
4.25 Trust Dimension
Care
4.00 Reliability
Integrity
3.75 Risk Taking
3.50
Friends Education Identity− Interest− Location−
 & Family  & Work  Based  Based  Based
Group Category
Figure 5.2: People have the highest trust in friends and family groups, and
lowest in interest- and location-based groups.
People use these groups more as places to transact and exchange (both physical
goods and information) than as places to build relationships [Granovetter, 1985].
By comparing groups across different categories, we can develop a more holistic
understanding of trust across different types of social groups that also draws on
insights from previously isolated studies [Moser et al., 2017, Holtz et al., 2017].
Activity
Here, we studied both a survey participant’s activity in a group as well as the
overall group activity across all members. Measures of activity include time
spent in the group and the number of actions (posts, likes, or comments) taken in
the group, averaged across the 28 days preceding the survey. In the case of public
groups, activity also included contributions from nonmembers. An individual’s
overall site engagement was not predictive of trust, and thus was excluded
from our analyses. Including activity features (time spent, group activity, and
participant in group activity) to the baseline model improves its adjusted R2 by
0.04 (p<.001).
149
Trust Score
As many activity features are correlated, we report coefficients when the
feature is independently added to the baseline model. Time spent in the group,
both by the individual (β=0.04, p<.001) and by other group members (β=0.05,
p<.001) independently predicts higher trust in groups. Overall, the number of
posts per member (β=0.07, p<.001), and the number of likes and comments per
post (β=0.07, p<.001) were also both independently associated with higher trust.
However, the number of comments and likes a participant made in a group was
associated with higher trust (β=0.10 on log scale, p<.001), but not the number of
posts the participant wrote.
Why is this the case? Posting in a group may be influenced by a variety of
factors other than trust (e.g., self-esteem [Forest and Wood, 2012]). In contrast,
likes and comments are forms of directed communication that people use to
maintain relationships with others [Ellison et al., 2014] and may therefore be
more conducive to building trust.
Homogeneity and Homophily
Trust may also be influenced by how similar people in a group are to each
other (homogeneity), and how similar an individual is to others in the group
(homophily).
For each group, we measured age and gender diversity by computing the
gender entropy of the group’s members and the standard deviation of their ages.
To measure homophily, we constructed a simple distance measure based on the
approach of [Abrahao et al., 2017]. If the participant had the same gender with
the majority of the group members, we coded the gender distance as 0, otherwise
150
1. If the participant’s age was within 5 years of the average age of the group,
we coded the age distance as 0, otherwise 1. The total distance from average
group members was calculated as the L1 distance, i.e., the sum of gender and age
distance ∈ (0,1,2). As different types of groups may have different demographic
compositions, we controlled for group category in this analysis.
Adding homogeneity and homophily features to the baseline model results in
a small improvement (increased adjusted R2 by less than 0.01, p<.001). Nonethe-
less, we found that both gender (β=0.04, p<.001) and age homogeneity (β=0.04,
p<.001) were associated with higher trust.
Surprisingly, homophily, measured as described above, was not predictive
of trust in groups. This contrasts with findings in previous work on trust and
homophily in dyadic exchange, which found that trust increases with gender
and age homophily [Abrahao et al., 2017, Ahmad et al., 2011]. While we only
studied age and gender homophily here, future work may consider other forms
of homophily (e.g., with respect to interests, location, or socio-economic status)
or other measures of homophily, especially in a group rather than dyadic context.
Network Structure
To understand how network structure mediates trust, we calculated the following
network features for each group: (1) network density: the number of friendships
in the entire group friendship graph divided by the number of possible combina-
tions; (2) average clustering coefficient: the average local clustering coefficient in
the group membership graph, which measures what proportion of an individ-
ual’s friends also know one another; (3) participant degree centrality: the number
151
(A) (B) (C) (D) (E)
Survey Participant Group Admin  Other Group Members
Figure 5.3: Groups differ in network density, participant degree centrality,
and how a participant’s friends are linked to each other. Each
node represents a group member. Each edge represents a friend-
ship between two members. The survey participant is colored
in red, and group admins are colored in yellow.
of friends a participant has in the group, normalized by group size; (4) k-core
existence: a measure of how a participant’s friends in the group are connected
with each other, calculated as whether a k-core component [Ugander et al., 2012]
exists for participant’s friendship graph in the group (we found that k=5 resulted
in the greatest model improvement); and (5) average mutual friend count: the mean
number of mutual friends between participant and group members.
Figure 5.3 illustrates how several group networks in our sample differ along
these network features. For example, Group A has higher network density and
higher average clustering coefficient than group B. Groups C and D differ in the
participant’s degree centrality. Group D contains a 5-core, but E does not.
These network features, when added to the baseline model, improves its
adjusted R2 by 0.10 (p<.001). Each feature was positively associated with trust
in groups (p<.001), though we note that these network features correlate highly
with one another. Considering these features separately, the average clustering
coefficient was most predictive (β=1.08, p<.001), followed by group density
(β=0.93, p<.001) and the participant’s degree centrality (β=0.84, p<.001).
152
Demographics 1.78%
General Attitudes Towards Others 20.63%
Basic Properties & Category 6.16%
Full Model
Activity 10.96% R 2= 0.26
Homogeneity 7.09%
Network Structure 22.04%
Figure 5.4: For each feature set, we calculated the average feature impor-
tance (measured by relative percent increase in MSE when a
feature is removed) in predicting trust in groups. Network
structure was the most important, followed by an individual’s
general attitudes towards others.
5.4.3 Predicting Trust in Groups
Thus far, we have shown how various sets of group characteristics separately
contribute to trust, after controlling for individual characteristics. Here, we
examine how these features can together predict composite trust in groups.
A random forest model that uses all feature sets (in both Table 5.3 and Ta-
ble 5.4) reached a performance of out-of-sample adjusted R2 of 0.26 and a mean-
squared error (MSE) of 0.53. We obtained similar performance using multiple
linear regression.
To understand the relative importance of the different feature sets, we ranked
all features by how much a random permutation of their values increased the
model’s MSE. These values are shown in Figure 5.4. We find that network features
are most important, followed by an individual’s general attitude towards others.
Least important were demographic features. Overall, this result suggests that
both individual and group characteristics are important in predicting trust in
153
groups.
Predicting Trust Using Only Observational Data
As we demonstrate relatively robust performance in predicting trust in groups,
one might consider using such predictions to make better group or community
recommendations. However, our model uses survey responses about individual
differences, including disposition to trust and related concepts, to make predic-
tions about trust in a specific group. In practical settings, it may not be feasible
to administer the survey questions on individual differences to all users. This
motivates the question of how well our modeling approach works in the absence
of the individual differences survey features. Excluding these features, our best
model obtained an adjusted R2 of 0.15 and MSE of 0.59. In this model, network
structure features were again most important, but instead followed by group
activity features.
5.4.4 Group Outcomes
Theoretical accounts of trust emphasize the impact of trust on community out-
comes, attributing trust to prosperity [Fukuyama, 1995], among other things.
Here, we analyze how trust relates to three different group outcomes: (1) the
percentage change in group size; (2) the percentage change in new tie formation
(the number of new ties divided by the number of pre-existing ties) among other
members of the group; and (3) the percentage change in new tie formation by
the survey participant in the group. All three measures were calculated by com-
paring the state of the group on the day of the survey to that 28 days after. As
154
Member+ > 1% Other Ties+ > 1% Participant Ties+ > 1%
80%
● ●
●
●
60%
●
40% ●
● ● ●
●
20%
● ●
● ●
0% ●
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Trust in Groups Composite
Figure 5.5: Groups with higher trust ratings are less likely to increase in
size (left), more likely for the survey participant to form new
connections in them (right), and had no effect on the likelihood
on forming friendships among group members other than the
rater (center).
these changes tend to be small, with a median change of about 1%, we instead
predict whether each measure would increase by more than 1%.
Figure 5.5 shows the percentage of groups that exhibited an increase by more
than 1% in each of the group outcomes listed above. Using logistic regression
and controlling for basic group properties such as group size, we found that
higher trust was associated with a lower likelihood of a group increasing in size
(odds ratio -0.87, p<.001); and a higher likelihood that the survey participant
would form more new friendships in the group (odds ratio 1.29, p<.001). Trust
in a group was not predictive of the likelihood of other group members forming
friendships in the group.
These results suggest a tension between trust and growth for online groups.
Our findings are consistent with previous work on online communities that
found that “cliquishness” (or high triangle density) makes a community less
attractive to join and less likely to grow in size [Backstrom et al., 2006]. While
membership growth is an important indicator of success for online groups [Kraut
and Fiore, 2014], trust, partially elicited by small groups and exclusive member-
155
Pct Groups w/ Increase
ship [Moser et al., 2017], can limit group expansion (but nonetheless encourages
individuals to make new connections within the group). Future work can ex-
amine the relationship between trust and group longevity, as well as other
interaction dynamics such as forming sub-communities within the group.
5.5 Discussion
In this work, we present a framework for predicting an individual’s trust in one
of their social groups on Facebook. Combining a large and diverse survey with
behavioral and demographic data, we show that both individual characteristics
and group characteristics contribute substantially to trust. We are able to explain
a significant portion of an individual’s trust in groups (R2=0.26) as well as show
how trust relates to outcomes such as membership growth and the formation of
new within-group friendships.
This work builds on many previous studies of trust in groups by showing
how features previously studied in isolation may interact with each other and
how important these features are relative to each other. Beyond confirming
that both an individual’s general disposition to trust others [Ferguson and Pe-
terson, 2015] and a group’s size [Brewer, 1991] affect that individual’s trust in
a group, we further show that group size matters less to individuals with a
greater disposition to trust, and that an individual’s feelings of receiving social
support from others in general is actually more predictive of trust in groups
than their general disposition to trust. Apart from demonstrating that people do
trust smaller, more private groups, we show that among groups with more than
150 members, the effect of exclusive membership decreases. Where previous
156
work has suggested a relation between group connectivity and trust [Coleman,
1988, Yuki et al., 2005], we also demonstrate that network measures such as the
average clustering coefficient in a group are among the strongest predictors of
trust in a group. Our findings on how directed communication such as likes
and comments contributes to group trust corroborate similar observations in
qualitative studies [Moser et al., 2017].
Nonetheless, several null results suggest areas for future exploration. While
prior work suggests trust differs with sociodemographic factors [Pew, 2007],
we found that age and gender explain close to zero variance in one’s trust in
groups. Future work may consider exploring other factors such as geography
or socioeconomic status. Cultural differences may also play a significant role
in trust: prior work found that an indirect relationship between two people
was more likely to increase trust for Japanese than Americans [Yuki et al., 2005],
suggesting that network structure may be more predictive of trust among the
former. Though we found that gender- and age-homogenous groups were more
trusted, we also found no evidence that gender or age homophily predicts trust
in groups, in contrast to previous literature suggesting that relationships between
similar individuals tend to be more trusting [Abrahao et al., 2017, Ahmad et al.,
2011, McPherson et al., 2001]. Understanding the extent to which these findings
apply to specific situations — moms’ buy-and-sell groups on Facebook are known
to garner trust [Moser et al., 2017] — remains future work.
Future work may also involve investigating other potential correlates of trust
such as psychological safety [Edmondson et al., 2004] and belonging [Zhao et al.,
2012], as well as other outcomes of trust on online groups. For example, high
trust may lead to a greater willingness to attend an event, share (or believe)
157
information originating from within the group, or donate to a cause.
5.5.1 Design Implications
The work reported here has several potential implications for the design of online
communities.
We showed that certain types of actions (e.g., commenting and liking) are
more positively associated with trust than others (e.g., posting). This adds
nuance to previous findings that people have greater trust in communities in
which they are more active [Cartwright and Zander, 1953]. As such, platforms
could prioritize facilitating directed interactions among group members, for
example, expanding features to support polling, brainstorming, and collective
planning. At the same time, these findings may also inform the design of content
recommendation systems. If these findings indicate that directed communication
is a key signal of trust, then incorporating such signals of directed communication
may better ensure that people see more content from communities that they trust
more.
Consistent with prior work [Kraut and Fiore, 2014], we found that trust
grows with the number of group admins and decreases with group size. As
online communities grow, it may be beneficial for platforms to encourage groups
to recruit additional admins to maintain existing levels of trust.
Further, network properties of online communities such as the average clus-
tering coefficient are strong predictors of trust. Adding members that increase
the average clustering of the group may be beneficial both to new members and
158
to the group as a whole.
Given that trust in a group correlates with behavioral signals, with additional
research, platforms may also be able to provide a rough indicator of trust in
groups and how it may be changing over time.
Last, our findings suggest alternative strategies for recommending groups to
individuals. For instance, recommending smaller, less popular groups may not
only increase the diversity of group recommendations, but also lead to greater
trust and user satisfaction.
5.5.2 Limitations
Our analysis is limited to groups on Facebook. Understanding how trust differs
in communities with different policies on anonymity (e.g., Reddit or Nextdoor)
or that have different feedback mechanisms remains an important area for future
exploration. Anonymity may increase trust by making it easier for vulnerable
populations to talk about sensitive issues, but also have a disinhibiting effect
and increase harassment and thus reduce trust [Kiesler et al., 1984]; indicators of
reputation or popularity such as up-votes and down-votes may also influence
trust, especially in the absence of other social signals [Resnick and Zeckhauser,
2002]. Still, many group properties (e.g., group size) that we examined apply
to groups in general; the interactions (e.g., posting or liking) that we looked at
are also common on other social media platforms. Along with the large number
and diversity of groups we surveyed, we expect that many of our findings will
generalize to other online communities3. While we controlled for individual
3Code to reproduce our analysis is available at https://github.com/
facebookresearch/trust-in-groups
159
differences such as demographics and an individual’s general attitudes towards
others, understanding differences that may arise in offline groups and with
regards to other factors such as location remains future work. Also, individuals
may choose to join groups based on other unobserved differences (e.g., word-
of-mouth). Finally, our methodology is based on correlations between variables
and cannot directly suggest causation. Most significantly, it is possible that
individuals who have different propensities to trust tend to join entirely different
groups, explaining some of our observed differences. Similar limitations apply
to the group outcomes analysis. While greater trust may lead one to connect to
other members of a group, it may also arise from making these connections.
5.6 Conclusion
Groups play a significant role in an individual’s social experiences and inter-
actions. Trust, which predicts numerous positive outcomes for a group and its
members, is core to a group’s proper functioning. In this work, we presented a
framework for predicting an individual’s trust in a social group, and identified
characteristics of both the individual and the group that help predict the indi-
vidual’s trust in the group. By surveying 6,383 Facebook Groups users about
their trust attitudes and examining aggregated behavioral and demographic
data for these individuals, we show that (1) an individual’s propensity to trust is
associated with how they trust their groups; (2) smaller, closed, older, more ex-
clusive, or more homogeneous groups are trusted more; and (3) a group’s overall
friendship-network structure and an individual’s position within that structure
can also predict trust. Last, we demonstrate how group trust predicts outcomes
at both individual and group level such as the formation of new friendship ties.
160
This work can contribute to future research and design decisions that better
support trust in online communities and foster long-term meaningful interactions
online and offline.
From a networked trust perspective, the work presented in this dissertation
shows that network structure is among the most predictive factors of trust in
groups. This highlights the importance of social networks in trust in networked
environment. Future work can investigate how social structure contributes
to trust in other platforms of digital exchange. For example, how does social
structure impact interpersonal trust in the context of hiring and professional
networking on LinkedIn?
At the same time, social structure can be important to understand other
phenomenon related to trust. Social groups have been found to be a vehicle for
misinformation [Silverman et al., 2018]. It is possible that trust in groups mediates
trust in information or misinformation shared to the group. Understanding how
social structure of the group impacts how people trust information shared in the
group can be an important direction for future work.
161
CHAPTER 6
DISCUSSION AND FUTURE WORK
6.1 Algorithms of Trust and Networked Trust
This dissertation has presented three different case studies of interpersonal trust
online using computational methods. In the first study, through a large-scale
analysis of user-generated image cues, I built algorithms that predict image
quality of products at a high accurately (87%). A controlled experiment further
showed that images selected by the algorithms outperform stock imagery in
generating perceived trustworthiness in the platform. Finally, image quality
predicted by the algorithm is correlated with higher sales in real world settings
on eBay.
In the second study, analysis of language cues in Airbnb host profiles uncov-
ered patterns of self-disclosure as well as how language cues lead to trust in
sharing economy platforms. For example, the topic of hospitality was most effec-
tive in establishing trust, controlling for profile length. These insights show the
importance of language cues and self-disclosure in establishing trust on sharing
economy platforms. Further, we were able to build algorithms to predict trust
in these profiles based on language cues (72% accuracy in recognizing profiles
that are perceived as high or low in trustworthiness for profiles less than twenty
words).
Finally, in the last study of trust, through a large-scale comprehensive study
of diverse Facebook social groups, I showed that the network features are among
the most predictive features of trust in social groups. People trust smaller, denser,
162
and more private groups. Similarly, I also developed algorithms to predict trust
in social groups based on a variety individual-level and group-level features.
Such algorithms can inform the ranking and recommendation for groups on
Facebook to promote trust.
Taken together, these three case studies of trust have extended our under-
standing of interpersonal trust online in different contexts: peer-to-peer mar-
ketplaces, sharing economy platforms, and social networks. These platforms
facilitate digitalized social exchange. As these platforms keep evolving, new
questions about trust will continue to be raised. At the same time, there are many
other platforms that facilitate social exchange and can be in scope for future in-
vestigation of trust: for example, video-based social network TikTok, Augmented
Reality (AR) enabled Snapchat, etc. Trust in these platforms is likely to also play
a very important role but future work is needed to understand specific contexts
and risks involved.
This dissertation also argues that the term “online trust” does not sufficiently
capture the factors that impact trust in networked environments. As people
are embedded in social and information networks when they conduct social
exchange, and as algorithms increasingly mediate such exchanges, a “networked
trust” view is proposed. Details on networked trust are laid out in Section 2.3.2.
As a reminder, networked trust has three focuses: (1) cues in Computer-Mediated
Communication; (2) embeddedness in social networks; and (3) increasing media-
tion by algorithms. Networked trust can be a useful framework for setting future
research agendas, which we discuss below.
163
6.2 Future Research Agenda
Here I discuss three areas of future research based on the work presented in this
dissertation: (1) networked trust and misinformation; (2) AI-Mediated Commu-
nication (AI-MC); and (3) AI-Mediated Exchange Theory (AI-MET).
6.2.1 Networked Trust and Misinformation
One of the most direct applications of the networked trust framework is in
understanding trust in information in misinformation research. In recent years,
concerns grew over the spread of misinformation online, especially through
social media [Allcott and Gentzkow, 2017, Allcott et al., 2019, Flintham et al.,
2018, Grinberg et al., 2019, Flintham et al., 2018]. Although misinformation,
defined as information that is false or misleading, has long existed before digital
platforms, it creates new challenges now as digital platforms and algorithms
optimized for engagement can result in viral diffusion of misinformation in social
networks [Zhang et al., 2018]. Prior work has begun to design solutions to curb
the spread of misinformation and achieved various levels of success [Zhang et al.,
2018].
Using the networked trust framework, we can organize prior work on misin-
formation into three focuses: (1) cues: contextual trust indicators; (2) networks: the
analysis of how misinformation spread in networks; and (3) algorithms: the role of
algorithms in configuring how misinformation spreads.
The first focus, contextual trust indicators, are cues signaling the trustworthi-
ness of information source [Jakesch et al., 2019,Pennycook and Rand, 2019,Zhang
164
et al., 2018, Kohring and Matthes, 2007]. For example, the Trust Project has de-
signed a core set of eight Trust Indicators, including factors around the historical
trustworthiness of the news outlet, the expertise of the author/reporter, and
the type of work (advertisement, opinion, or news reporting)1. The trust indi-
cators often rely on a crowdsourcing approach, either through experts [Zhang
et al., 2018] or through laypeople, which has been shown to be successful in
discriminating high and low quality content and provide ratings that highly
correlate with professional checkers [Pennycook and Rand, 2019, Epstein et al.,
2019]. Relatedly, prior work has examined how people evaluate news credibility
based on source, trust indicators, and related factors such as their own political
beliefs and news literacy [Tully et al., 2019,Sundar, 1998, Jakesch et al., 2019]. On
Twitter, verified status of the news distributor was not shown to have an effect
on credibility of the news being shared [Vaidya et al., 2019]. Trust indicators act
as cues to signal the trustworthiness of another party (usually the news source),
and news consumers subsequently make decisions about trust based on these
cues while considering their own propensity to trust and other factors.
The second focus, understanding how misinformation spread in networks,
aligns well with the networks focus in networked trust framework. Early work
on information diffusion has established that those who are exposed to a social
feed in social networks are significantly more likely to spread information, and
do so sooner than those who are not exposed [Bakshy et al., 2012]. In addition,
weak ties play a more dominant role in the dissemination of information online,
compared to strong ties [Bakshy et al., 2012]. Work on misinformation has
also established patterns on how true and false information travel through
networks [Vosoughi et al., 2018, Budak, 2019]. For example, false news were
1https://thetrustproject.org/faq/indicator
165
found to diffuse “significantly farther, faster, deeper, and more broadly than the
truth in all categories of information” [Vosoughi et al., 2018]. In addition, social
bots were found to play a key role in the spread of low-credibility content [Shao
et al., 2018]. There has also been qualitative work focused on understanding how
people consume news in the context of social networks, highlighting different
strategies people use to establish perceptions of trust while being embedded
in networks. Although people rely on source and content to evaluate news
credibility in social networks, interests in the topic also play a very important
role — “when the topic of the story was not seen as personally relevant, there was
little interest in figuring out whether or not it was true” [Flintham et al., 2018]. In
addition, research has shown that when news is shared in social networks, trust
indicators might be ignored, especially when they come from strangers rather
than pre-existing social relationships [Hannak et al., 2014]. The networks focus
of the above-mentioned misinformation research aligns well with the second
focus of networked trust, which views interpersonal trust not in isolated dyadic
interactions, but rather embedded in social and information networks where the
structure of the network play an important role.
Finally, the algorithms focus investigates the role of algorithms in configuring
how misinformation spread. As research on the algorithmic transparency, bias,
and accountability gains wider interest [Barocas and Selbst, 2016], algorithmic
transparency in the context of news media is also receiving more attention [Di-
akopoulos and Koliska, 2017]. Given that the majority of social media has have
heavily algorithmic curated feeds [Eslami et al., 2015, Eslami et al., 2019], there
is growing concern over the opacity of algorithms, especially around how such
opacity affects news consumption [Diakopoulos and Koliska, 2017]. The role
of algorithms in perpetuating misinformation can be found outside of social
166
networks as well, for example, in Amazon book recommendations2 and Google
search3. This algorithms focus aligns well with the third aspect of networked
trust, which is concerned with understanding algorithmic mediation and how
people’s trust in algorithms might play an intermediating role on trust in news.
Based on this networked trust view, we can develop research questions to
deepen our understanding of trust in misinformation research, especially how
different focuses of networked trust interact. Instead of focusing on understand-
ing how cues, networks, and algorithms work in isolation, we can begin to ask
questions around how they interact or work together in the networked envi-
ronment. For example, social contexts extracted from social networks can help
improve the prediction of news credibility [Shu et al., 2019b, Shu et al., 2017, Shu
et al., 2018] compared to using language cues alone [Yao et al., 2017a, Rashkin
et al., 2017, Popat et al., 2016].
Below are some other examples of questions we can ask about trust in misin-
formation research. How does news source interact with social signals extracted
from networks about the distributor (e.g., from whom the news is being shared
and their social status)? How do cues (trust indicators) affect the diffusion of mis-
information in social networks? Building directly off my work on trust in social
groups [Ma et al., 2019a], how do people trust and interact with misinformation
differently when false information is shared in groups with different network
structures?
We can also ask how algorithms affect cues. How can we build personal-
ization algorithms that are perceived as more trustworthy [Lai and Tan, 2018]?
2https://www.wired.com/story/amazon-and-the-spread-of-health-misinformation/
3https://www.blog.google/around-the-globe/google-europe/fighting-disinformation-
across-our-products/
167
How can we build trustworthy algorithms that detect misinformation, perhaps
through high transparency and explainability [Shu et al., 2019a]? These questions
can help further advance our understanding of misinformation in networked
environments.
6.2.2 AI-Mediated Communication (AI-MC)
The second future research direction that builds off the work presented in this
dissertation is the area of AI-Mediated Communication (AI-MC). Recall that the
first focus of networked trust framework is investigating how cues in Computer-
Mediated Communication affect interpersonal outcomes. However, as discussed
in the end of Chapter 4, algorithms can modify, augment, and even generate cues
to optimize for perceived trustworthiness in online communication.
Such algorithmic ability to modify, augment, and generate cues in presen-
tation online raises new questions about interpersonal trust and other commu-
nication outcomes. Specifically, several real-world examples have shown that
AI-powered systems impact online interpersonal communication processes. For
example, Google’s Smart Reply, a system that generates short email responses,
has already been deployed in real-world systems and accounts for 10% of the
mobile replies in the Google Inbox [Kannan et al., 2016]. Other systems also medi-
ate interpersonal interactions, such as profile summary auto-generation feature
on LinkedIn4; a startup named CV Compiler that boasts using machine learning
to revise resumes to boost chances of getting a high-paying job in tech5; and chat
bots that suggests responses to other people’s messages mid-conversation [Ho-
4https://www.linkedin.com/pulse/solving-blank-slate-problem-through-auto-generated-
summary-jalan/
5https://cvcompiler.com/
168
henstein and Jung, 2018].
These AI-powered communication systems raise new questions about online
interpersonal communication, especially around interpersonal trust. The term
“Computer-Mediated” is no longer sufficient to capture the additional complexity
that AI systems bring in interpersonal exchange. Rather than simply being a pas-
sive mediator, the AI-powered systems may alter the messages being transmitted,
increasing uncertainty and risks in interpersonal communication.
To capture these new challenges brought by AI-powered communication
systems, I developed the term “AI-Mediated Communication” to chart an emerg-
ing area of research. At the time of this writing, the exact definition and re-
search questions represented by the term are still being developed [Naaman
et al., 2019]. However, first attempts have already established observed effects
of AI-Mediated Communication [Jakesch et al., 2019], by extending the work
on computational trustworthiness of Airbnb host profiles presented in Chap-
ter 4 [Ma et al., 2017a,Ma et al., 2017c]. Experiments show that there is a decrease
in perceived trustworthiness when people are not sure whether a profile is writ-
ten by human or AI [Jakesch et al., 2019]. In other words, the lack of transparency
in the existence of AI hinders interpersonal trust. Future work can continue
to investigate and clarify mental models around transparency in AI-Mediated
Communication. Another emerging line of work in the area of AI-Mediated
Communication examines how AI-suggested responses impact people’s behav-
ior. For example, there is early evidence that positivity bias in AI-suggested
responses might nudge people to respond differently to requests [Hohenstein
and Jung, 2018].
Future developments in AI-Mediated Communication might also address
169
other important questions: For example, what characteristics of the AI systems
that mediate interpersonal communication lead to acceptance and adoption?
What characteristics lead to distrust and avoidance? The insights from AI-
Mediated Communication will be valuable to further the emerging discussion
in CHI community around design guidelines in human-AI interaction [Amer-
shi et al., 2019]. Another question we can ask is how these AI systems impact
decision-making. Behavior science has uncovered how people systematically
made decisions that are not rational, or subject to priming and anchoring [Ariely,
2008, Kahneman, 2011]. Can AI-powered communication systems amplify or
curb such patterns in decision-making? Finally, hard ethical questions need to
be raised with regard to these systems, such as whether the algorithms used in
AI-powered communication systems are fair, transparent, and accurate. Chal-
lenges in misrepresentation and algorithmic manipulation might exacerbate the
problem of misinformation online [Susser et al., 2018]. As AI-Mediated Commu-
nication’s definition and research agenda continue to develop [Naaman et al.,
2019], AI-Mediated Communication will have important implications for the
design, implementation, as well as the regulation of AI-powered systems that
mediate interpersonal communication online.
6.2.3 AI-Mediated Exchange Theory (AI-MET)
In the definition of AI-Mediated Communication, one of the boundary condition
is that “whether a computational agent is operating on behalf of a communica-
tor engaged in an interpersonal interaction” [Naaman et al., 2019]. Under this
condition, algorithms that perform the ranking, recommendations, and classifi-
cations that support human communication are not included for consideration.
170
Although this boundary condition is useful for sharpening the definition and
research agenda for AI-Mediated Communication, the exclusion of other types
of algorithms, especially ranking and personalization algorithms, creates gaps in
understanding the role of AI in digitalized exchange.
The inherent assumption of AI-Mediated Communication is that two parties
are engaging in a direct exchange relationship through communication online.
However, as I reviewed in Chapter 2, social exchange theory, the sociological
theory that this dissertation builds the definition of trust on, states that there can
be both direct and generalized (indirect) exchanges. Direct exchange occurs when
there is a transfer of resources or information between two parties. In generalized
exchange, however, often people pool resources together collectively to create
greater values, and then re-distribute the values among the group [Yamagishi and
Cook, 1993, Bearman, 1997]. In the digital context, open source communities and
peer-to-peer platforms are considered as examples of generalized information
exchange [Cheshire, 2007].
It is important, then, to consider the effect of algorithms in mediating general-
ized exchange relationships in addition to direct exchange. Future work can work
on developing “AI-Mediated Exchange Theory (AI-MET)” as a framework to
consider the role of AI in mediating social exchange relationships.
Specifically, the AI-Mediated Exchange Theory can discuss different mecha-
nisms of algorithmic mediation. Defining different mechanisms through which
algorithms mediate exchange can be helpful in organizing current and future
work around algorithms’ impact on social interactions and society in general,
especially around trust. Examples of such algorithmic mediation mechanisms
might include: algorithmic aggregation, algorithmic representation, algorithmic
171
curation, and algorithmic augmentation.
1. Algorithmic aggregation can refer to the process where algorithms summa-
rize information that is available about a potential exchange partner and
make the aggregate information available to people on exchange networks.
Examples of algorithmic aggregation include, reputation systems, social
signals such as the number of likes, comments, and up- or down-voting.
2. Algorithmic representation can refer to the process where algorithms try to
represent a specific entity using a label or a high dimensional vector. Ex-
amples of algorithmic representation include clustering algorithms, word
embeddings, and topic modeling. Classification algorithms such as facial
detection [Buolamwini and Gebru, 2018] can also be considered as exam-
ples of algorithmic representation as they represent the input data with a
label (1-dimensional vector). These representations are then fed into other
algorithms, such as recommendation systems, to further affect how people
experience the digital platforms.
3. Algorithmic curation can refer to the process where algorithms select, or-
ganize and present information based on specific goals. Examples of the
algorithmic curation include, news feed ranking [Eslami et al., 2015,Eslami
et al., 2019], personalization and content recommendation algorithms, as
well as algorithms that assist decision-making such in Applicant Tracking
Systems (ATS) in hiring [Mukherjee et al., 2014,Chiang and Suen, 2015,Bald-
win et al., 2014, Arya et al., 2015].
4. Finally, algorithmic augmentation can refer to the process where algorithms
modify, augment, and even generate information related to exchange.
This mechanism is essentially the focus of AI-Mediated Communication.
172
Examples of algorithmic augmentation include, deep fake [Bansal et al.,
2018,Chan et al., 2018] for images and text, and the ability to generate realis-
tic writings and images based on Generative Adversarial Nets [Goodfellow
et al., 2014, Karras et al., 2018, Wang, 2019].
AI-Mediated Exchange Theory provides the possibility to understand how
different algorithmic mediation mechanisms fit together in real-world systems to
influence digitalized exchange. For example, in the hiring context, AI-Mediated
Communication will be most concerned about AI augmentation — how algo-
rithms that can improve resumes, or to generate resumes altogether might impact
interpersonal outcomes in hiring. However, other algorithms such as the systems
used for ranking and filtering participants can have inherent biases (e.g., gender
bias [Gee, 2017]) and can affect interpersonal outcomes through AI curation.
AI-MET allows for the discussion of how different mechanisms cascade to affect
interpersonal outcomes. Here, we can examine how AI augmentation interacts
with AI curation in the context of fairness in hiring.
In another example, in the discussion of misinformation online, AI-Mediated
Communication may primarily focuses on AI augmentation — understanding
how people perceive others differently in the presence of AI that can generate re-
alistic fake videos. However, as discussed above, one of the key areas of research
to understand misinformation in networked environment is through understand-
ing how misinformation spread in social networks, and the algorithms’ role in
recommending and amplifying misinformation in networks (AI curation). AI
augmentation and curation again can cascade to affect trust in information in the
context of misinformation. Again, under AI-MET, different algorithms be studied
in relation to one another to understand how different algorithmic mediation
173
mechanisms impact interpersonal social exchange and outcomes.
In summary, three future research agenda were charted based on the work
presented in this dissertation: (1) networked trust and misinformation; (2) AI-
Mediated Communication (AI-MC); and (3) AI-Mediated Exchange Theory (AI-
MET). These research directions either directly apply the networked trust lens
in new contexts (misinformation), or extend and expand specific areas of focus
of the networked trust. AI-Mediated Communication extends the first focus of
networked trust, cues in Computer-Mediated Communication; and AI-Mediated
Exchange Theory expands the last focus of networked trust: algorithmic medi-
ation. These research directions will continue to address important questions
about trust in the networked environment.
6.3 Conclusion
In conclusion, this dissertation presents a series of work on understanding
interpersonal trust computationally, while proposing the new framework of
“networked trust”. Three work presented in this dissertation detail three different
computational framework on understanding and predicting trust using image
cues, language cues, and social networks, in the contexts of online commerce,
sharing economy, and social network groups respectively. The computational
approaches help us not only gain a deeper understanding of trust in each context,
but also can inform the design of digital platforms that facilitate exchange to
promote trust. The algorithms of trust developed can also directly be deployed
in these systems to rank and recommend exchange partners that are likely to be
perceived as trustworthy to increase the overall trust on platforms.
174
Further, I also developed the framework of “networked trust” — a new frame-
work for thinking about trust in hyper-connected digital environments with three
focuses: (1) cues in Computer-Mediated Communication; (2) embeddedness in
social networks; and (3) increasing mediation by algorithms. As digital platforms
continue to mediate social exchange in new ways, changes to social exchange
structure will continue to bring new challenges and questions about interpersonal
trust in networked environments. The networked trust view can be generalized
to other contexts, for example, misinformation, to help organize prior work and
chart new research directions on trust in misinformation research. Other con-
texts where networked trust can be useful include online creative marketplaces
(e.g., Spotify, Artsy, Patreon), crowdfunding platforms (e.g., Kiva, Kickstarter),
and other sharing economy offerings (e.g., Airbnb Experiences). In addition,
extensions of networked trust can lead to promising new areas of research, such
as AI-Mediated Communication (AI-MC) and AI-Mediated Exchange Theory
(AI-MET). Initial ideas for both AI-MC and AI-MET are laid out above. Future
work can develop these ideas further into fuller studies and theories.
Importantly, one key limitation of this dissertation is that the flip side of trust,
including distrust, bias and discrimination, has not been adequately addressed,
especially in networked settings. Future work is needed to understand how
digital mediation can lead to the opposite of trust, and how it leads to issues in
algorithmic fairness, accountability, and transparency (FAT*). In particular, future
work is needed to understand the fairness, transparency, and explainability of
the algorithms predicting trust, before they are deployed in real-world systems.
Finally, to highlight how much trust has changed in the past two decades,
I give one last example of social exchange that requires trust — taxi drivers.
175
One of the foundational book on trust, Streetwise, by Gambetta and Hamill and
published in 2005, investigated how taxi drivers assess the trustworthiness of
prospective passengers in Belfast, Northen Ireland, and New York. Through
the novel use of signaling theory, they show the mechanisms taxi drivers use to
establish trust rely heavily on physical cues — for example, picking up passen-
gers only at well-lit corners and not in dark alleys [Gambetta and Hamill, 2005].
However, a little over a decade later, on Uber or Lyft, such decisions about which
passengers to pick up are largely made online, through virtual networks, and
increasingly controlled by algorithms [Lee et al., 2015b,Rosenblat, 2018,Rosenblat
and Stark, 2016]. Real-world physical cues are discarded, while cues, networks,
and algorithms become more and more important. This shift is how networked
trust came into being with the digitalization of social exchange.
If anything, we are still in early stages of yet another wave of change in the
digitalization of social exchange, where algorithms gain increasing importance.
This dissertation has set the foundation for understanding how algorithms me-
diate social exchange on digital platforms. Future research can further develop
AI-Mediated Exchange Theory to better understand the long-term consequences
of algorithms on interpersonal outcomes in exchange, especially trust.
176
BIBLIOGRAPHY
[Abrahao et al., 2017] Abrahao, B., Parigi, P., Gupta, A., and Cook, K. S. (2017).
Reputation offsets trust judgments based on social biases among airbnb users.
Proceedings of the National Academy of Sciences, 114(37):9848–9853.
[Adamic et al., 2008] Adamic, L. A., Zhang, J., Bakshy, E., and Ackerman, M. S.
(2008). Knowledge sharing and yahoo answers: everyone knows something.
In Proceedings of the International Conference on World Wide Web, pages 665–674.
ACM.
[Ahmad et al., 2011] Ahmad, M. A., Ahmed, I., Srivastava, J., and Poole, M. S.
(2011). Trust me, i’m an expert: Trust, homophily and expertise in mmos. In
Proceedings of the IEEE International Conference on Privacy, Security, Risk and
Trust, pages 882–887. IEEE.
[Ainsworth et al., 2015] Ainsworth, M. D., Blehar, M. C., Waters, E., and Wall,
S. N. (2015). Patterns of Attachment: A Psychological Study of the Strange Situation.
Psychology Press.
[Airbnb, 2019] Airbnb (2019). What is a superhost? https://www.airbnb.
com/help/article/828/what-is-a-superhost. (Accessed Jul 2019).
[Akerlof, 1978] Akerlof, G. A. (1978). The market for “lemons”: Quality uncer-
tainty and the market mechanism. In Uncertainty in Economics, pages 235–251.
Elsevier.
[Allcott and Gentzkow, 2017] Allcott, H. and Gentzkow, M. (2017). Social media
and fake news in the 2016 election. Journal of Economic Perspectives, 31(2):211–
36.
[Allcott et al., 2019] Allcott, H., Gentzkow, M., and Yu, C. (2019). Trends in the
diffusion of misinformation on social media. Technical report, National Bureau
of Economic Research.
[Amershi et al., 2019] Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi,
B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., et al. (2019).
Guidelines for human-ai interaction. In Proceedings of the ACM Conference on
Human Factors in Computing Systems.
[Antin et al., 2011] Antin, J., Yee, R., Cheshire, C., and Nov, O. (2011). Gender
177
differences in wikipedia editing. In Proceedings of the International Symposium
on Wikis and Open Collaboration, pages 11–14. ACM.
[Ariely, 2008] Ariely, D. (2008). Predictably Irrational. Harper Audio.
[Arrow, 1974] Arrow, K. J. (1974). The Limits of Organization. WW Norton &
Company.
[Arya et al., 2015] Arya, D., Ha-Thuc, V., and Sinha, S. (2015). Personalized
federated search at linkedin. In Proceedings of the ACM International Conference
on Information and Knowledge Management, pages 1699–1702. ACM.
[Ashleigh et al., 2012] Ashleigh, M. J., Higgs, M., and Dulewicz, V. (2012). A
new propensity to trust scale and its relationship with individual well-being:
implications for hrm policies and practices. Human Resource Management
Journal, 22(4):360–376.
[Bachmann, 2001] Bachmann, R. (2001). Trust, power and control in trans-
organizational relations. Organization Studies, 22(2):337–365.
[Backstrom et al., 2006] Backstrom, L., Huttenlocher, D., Kleinberg, J., and Lan,
X. (2006). Group formation in large social networks: membership, growth,
and evolution. In Proceedings of the ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pages 44–54. ACM.
[Backstrom and Kleinberg, 2014] Backstrom, L. and Kleinberg, J. (2014). Ro-
mantic partnerships and the dispersion of social ties: a network analysis
of relationship status on facebook. In Proceedings of the ACM Conference on
Computer-Supported Cooperative Work & Social Computing, pages 831–841. ACM.
[Bakshy et al., 2015] Bakshy, E., Messing, S., and Adamic, L. A. (2015). Exposure
to ideologically diverse news and opinion on facebook. Science, 348(6239):1130–
1132.
[Bakshy et al., 2012] Bakshy, E., Rosenn, I., Marlow, C., and Adamic, L. (2012).
The role of social networks in information diffusion. In Proceedings of the
International Conference on World Wide Web, pages 519–528. ACM.
[Baldwin et al., 2014] Baldwin, T., Chang, C., VanGeest, J. R., and Derezin, M.
(2014). Techniques for using social proximity scores in recruiting and/or
hiring. US Patent App. 14/015,751.
178
[Bansal et al., 2018] Bansal, A., Ma, S., Ramanan, D., and Sheikh, Y. (2018).
Recycle-gan: Unsupervised video retargeting. In Proceedings of the European
Conference on Computer Vision, pages 119–135.
[Barocas and Selbst, 2016] Barocas, S. and Selbst, A. D. (2016). Big data’s dis-
parate impact. California Law Review, 104:671.
[Barrera Jr and Ainlay, 1983] Barrera Jr, M. and Ainlay, S. L. (1983). The structure
of social support: A conceptual and empirical analysis. Journal of Community
Psychology, 11(2):133–143.
[Bazarova and Choi, 2014] Bazarova, N. N. and Choi, Y. H. (2014). Self-
disclosure in social media: Extending the functional approach to disclosure
motivations and characteristics on social network sites. Journal of Communica-
tion, 64(4):635–657.
[Bearman, 1997] Bearman, P. (1997). Generalized exchange. American Journal of
Sociology, 102(5):1383–1415.
[Berg et al., 1995] Berg, J., Dickhaut, J., and McCabe, K. (1995). Trust, reciprocity,
and social history. Games and Economic Behavior, 10(1):122–142.
[Berger and Calabrese, 1975] Berger, C. R. and Calabrese, R. J. (1975). Some
explorations in initial interaction and beyond: Toward a developmental theory
of interpersonal communication. Human Communication Research, 1(2):99–112.
[Bhattacharya et al., 2010] Bhattacharya, S., Sukthankar, R., and Shah, M. (2010).
A framework for photo-quality assessment and enhancement based on visual
aesthetics. In Proceedings of the ACM International Conference on Multimedia,
pages 271–280. ACM.
[Bijlsma and Koopman, 2003] Bijlsma, K. and Koopman, P. (2003). Introduction:
trust within organisations. Personnel Review, 32(5):543–555.
[Bjørnskov, 2007] Bjørnskov, C. (2007). Determinants of generalized trust: A
cross-country comparison. Public Choice, 130(1-2):1–21.
[Bland et al., 2007] Bland, E. M., Black, G. S., and Lawrimore, K. (2007). Risk-
reducing and risk-enhancing factors impacting online auction outcomes: em-
pirical evidence from ebay auctions. Electronic Commerce Research, 8(4):236.
[Blau and Kahn, 2007] Blau, F. D. and Kahn, L. M. (2007). The gender pay gap:
179
Have women gone as far as they can? Academy of Management Perspectives,
21(1):7–23.
[Blau, 1964] Blau, P. (1964). Exchange and Power in Social Life. John Wiley & Sons.
[Bolukbasi et al., 2016] Bolukbasi, T., Chang, K.-W., Zou, J. Y., Saligrama, V.,
and Kalai, A. T. (2016). Man is to computer programmer as woman is to
homemaker? debiasing word embeddings. In Advances in Neural Information
Processing Systems, pages 4349–4357.
[Boss, 1978] Boss, R. W. (1978). Trust and managerial problem solving revisited.
Group & Organization Studies, 3(3):331–342.
[Bowlby, 1969] Bowlby, J. (1969). Attachment and Loss: Attachment. Basic books.
[boyd and Ellison, 2007] boyd, d. and Ellison, N. (2007). Social network sites:
Definition, history, and scholarship. Journal of Computer-Mediated Communica-
tion, 13(1):210–230.
[Boyd and Ellison, 2007] Boyd, D. M. and Ellison, N. B. (2007). Social network
sites: Definition, history, and scholarship. Journal of Computer-Mediated Com-
munication, 13(1):210–230.
[Bradley and Terry, 1952] Bradley, R. A. and Terry, M. E. (1952). Rank analysis
of incomplete block designs: I. the method of paired comparisons. Biometrika,
39(3/4):324–345.
[Brewer, 1991] Brewer, M. B. (1991). The social self: On being the same and
different at the same time. Personality and Social Psychology Bulletin, 17(5):475–
482.
[Budak, 2019] Budak, C. (2019). What happened? the spread of fake news
publisher content during the 2016 us presidential election. In Proceedings of the
International Conference on World Wide Web, pages 139–150. ACM.
[Buolamwini and Gebru, 2018] Buolamwini, J. and Gebru, T. (2018). Gender
shades: Intersectional accuracy disparities in commercial gender classification.
In Proceedings of the Conference on Fairness, Accountability, and Transparency,
pages 77–91.
[Butler Jr, 1999] Butler Jr, J. K. (1999). Trust expectations, information sharing,
180
climate of trust, and negotiation effectiveness and efficiency. Group & Organi-
zation Management, 24(2):217–238.
[Cartwright and Zander, 1953] Cartwright, D. and Zander, A. (1953). Group
cohesiveness: introduction. Group Dynamics: Research and Theory. Evanston, IL:
Row Peterson.
[Chamberlain, 2016] Chamberlain, A. (2016). Demystifying the gender pay gap.
Mill Valley, CA: Glassdoor.
[Chan et al., 2018] Chan, C., Ginosar, S., Zhou, T., and Efros, A. A. (2018). Every-
body dance now. arXiv preprint arXiv:1808.07371.
[Chanley et al., 2000] Chanley, V. A., Rudolph, T. J., and Rahn, W. M. (2000). The
origins and consequences of public trust in government: A time series analysis.
Public Opinion Quarterly, 64(3):239–256.
[Chaudhuri and Holbrook, 2001] Chaudhuri, A. and Holbrook, M. B. (2001).
The chain of effects from brand trust and brand affect to brand performance:
the role of brand loyalty. Journal of Marketing, 65(2):81–93.
[Cheng et al., 2014] Cheng, J., Adamic, L., Dow, P. A., Kleinberg, J. M., and
Leskovec, J. (2014). Can cascades be predicted? In Proceedings of the Interna-
tional Conference on World Wide Web, pages 925–936. ACM.
[Cheshire, 2007] Cheshire, C. (2007). Selective incentives and generalized infor-
mation exchange. Social Psychology Quarterly, 70(1):82–100.
[Cheshire, 2011] Cheshire, C. (2011). Online trust, trustworthiness, or assurance?
Daedalus, 140(4):49–58.
[Cheshire and Cook, 2004] Cheshire, C. and Cook, K. S. (2004). The emergence
of trust networks under uncertainty-implications for internet interactions.
Analyse und Kritik, 26(1):220.
[Chiang and Suen, 2015] Chiang, J. K.-H. and Suen, H.-Y. (2015). Self-
presentation and hiring recommendations in online communities: Lessons
from linkedin. Computers in Human Behavior, 48:516–524.
[Choi and Mai, 2018] Choi, Y. and Mai, D. Q. (2018). The sustainable role of the
e-trust in the b2c e-commerce of vietnam. Sustainability, 10(1).
181
[Chung et al., 2012] Chung, S. H., Goswami, A., Lee, H., and Hu, J. (2012). The
impact of images on user clicks in product search. In Proceedings of the Interna-
tional Workshop on Multimedia Data Mining, pages 25–33. ACM.
[Coleman, 1988] Coleman, J. S. (1988). Social capital in the creation of human
capital. American Journal of Sociology, 94:S95–S120.
[Colquitt et al., 2011] Colquitt, J. A., LePine, J. A., Zapata, C. P., and Wild, R. E.
(2011). Trust in typical and high-reliability contexts: Building and reacting to
trust among firefighters. Academy of Management Journal, 54(5):999–1015.
[Colquitt et al., 2007] Colquitt, J. A., Scott, B. A., and LePine, J. A. (2007). Trust,
trustworthiness, and trust propensity: a meta-analytic test of their unique rela-
tionships with risk taking and job performance. Journal of Applied Psychology,
92(4):909.
[Cook et al., 2009] Cook, K. S., Snijders, C., Buskens, V., and Cheshire, C. (2009).
eTrust: Forming Relationships in the Online World: Forming Relationships in the
Online World. Russell Sage Foundation.
[Cook and Whitmeyer, 1992] Cook, K. S. and Whitmeyer, J. M. (1992). Two
approaches to social structure: Exchange theory and network analysis. Annual
Review of Sociology, 18(1):109–127.
[Cook et al., 2005] Cook, K. S., Yamagishi, T., Cheshire, C., Cooper, R., Matsuda,
M., and Mashima, R. (2005). Trust building via risk taking: A cross-societal
experiment. Social Psychology Quarterly, 68(2):121–142.
[Corbitt et al., 2003] Corbitt, B. J., Thanasankit, T., and Yi, H. (2003). Trust and
e-commerce: a study of consumer perceptions. Electronic Commerce Research
and Applications, 2(3):203–215.
[Corritore et al., 2001] Corritore, C. L., Wiedenbeck, S., and Kracher, B. (2001).
The elements of online trust. In Companion to the ACM Conference on Human
Factors in Computing Systems, pages 504–505. ACM.
[Costa et al., 2001] Costa, A. C., Roe, R. A., and Taillieu, T. (2001). Trust within
teams: The relation with performance effectiveness. European Journal of Work
and Organizational Psychology, 10(3):225–244.
[Cozby, 1972] Cozby, P. C. (1972). Self-disclosure, reciprocity and liking. Sociom-
etry, pages 151–160.
182
[Danescu-Niculescu-Mizil et al., 2013a] Danescu-Niculescu-Mizil, C., Sudhof,
M., Jurafsky, D., Leskovec, J., and Potts, C. (2013a). A computational ap-
proach to politeness with application to social factors. In Proceedings of the
Annual Meeting of the Association for Computational Linguistics.
[Danescu-Niculescu-Mizil et al., 2013b] Danescu-Niculescu-Mizil, C., Sudhof,
M., Jurafsky, D., Leskovec, J., and Potts, C. (2013b). A computational approach
to politeness with application to social factors. In Proceedings of the Annual
Meeting of the Association for Computational Linguistics.
[Datta et al., 2018] Datta, A., Makagon, J., Mulligan, D., and Tschantz, M. (2018).
Discrimination in online advertising a multidisciplinary inquiry. In Proceedings
of the Conference on Fairness, Accountability, and Transparency.
[Datta et al., 2006] Datta, R., Joshi, D., Li, J., and Wang, J. Z. (2006). Studying aes-
thetics in photographic images using a computational approach. In Proceedings
of the European Conference on Computer Vision, pages 288–301. Springer.
[Delgado-Ballester and Luis Munuera-Alemán, 2001] Delgado-Ballester, E. and
Luis Munuera-Alemán, J. (2001). Brand trust in the context of consumer
loyalty. European Journal of Marketing, 35(11/12):1238–1258.
[Deng et al., 2009] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L.
(2009). Imagenet: A large-scale hierarchical image database. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition. IEEE.
[Denson et al., 2006] Denson, T. F., Lickel, B., Curtis, M., Stenstrom, D. M., and
Ames, D. R. (2006). The roles of entitativity and essentiality in judgments of
collective responsibility. Group Processes & Intergroup Relations, 9(1):43–61.
[Denters, 2002] Denters, B. (2002). Size and political trust: evidence from den-
mark, the netherlands, norway, and the united kingdom. Environment and
Planning C: Government and Policy, 20(6):793–812.
[Derlega et al., 1987] Derlega, V. J., Winstead, B. A., Wong, P., and Greenspan,
M. (1987). Self-disclosure and relationship development: An attributional
analysis. Interpersonal Processes: New Directions in Communication Research,
pages 172–187.
[DeSteno et al., 2012] DeSteno, D., Breazeal, C., Frank, R. H., Pizarro, D., Bau-
mann, J., Dickens, L., and Lee, J. J. (2012). Detecting the trustworthiness of
novel partners in economic exchange. Psychological Science, 23(12):1549–1556.
183
[Di et al., 2014] Di, W., Sundaresan, N., Piramuthu, R., and Bhardwaj, A. (2014).
Is a picture really worth a thousand words? on the role of images in e-
commerce. In Proceedings of the ACM International Conference on Web Search and
Data Mining, pages 633–642. ACM.
[Diakopoulos and Koliska, 2017] Diakopoulos, N. and Koliska, M. (2017). Algo-
rithmic transparency in the news media. Digital Journalism, 5(7):809–828.
[Dirks, 1999] Dirks, K. T. (1999). The effects of interpersonal trust on work group
performance. Journal of Applied Psychology, 84(3):445.
[Dirks and Ferrin, 2002] Dirks, K. T. and Ferrin, D. L. (2002). Trust in leadership:
Meta-analytic findings and implications for research and practice. Journal of
Applied Psychology, 87(4):611.
[Doleac and Stein, 2013] Doleac, J. L. and Stein, L. C. (2013). The visible hand:
Race and online market outcomes. The Economic Journal, 123(572).
[Donath, 2007] Donath, J. (2007). Signals in social supernets. Journal of Computer-
Mediated Communication, 13(1):231–251.
[Dubey et al., 2017] Dubey, A., Abhinav, K., Hamilton, M., and Kass, A. (2017).
Analyzing gender pay gap in freelancing marketplace. In Proceedings of the
ACM Conference on Computers and People Research, pages 13–19. ACM.
[Dunbar, 1992] Dunbar, R. I. (1992). Neocortex size as a constraint on group size
in primates. Journal of Human Evolution, 22(6):469–493.
[Dwyer et al., 2007] Dwyer, C., Hiltz, S., and Passerini, K. (2007). Trust and
privacy concern within social networking sites: A comparison of facebook and
myspace. Proceedings of the Americas Conference on Information Systems, page
339.
[Easley et al., 2010] Easley, D., Kleinberg, J., et al. (2010). Networks, Crowds, and
Markets, volume 8. Cambridge University Press.
[eBay, 2019] eBay (2019). Photo tips. https://pages.ebay.com/
seller-center/listing-and-marketing/photo-tips.html. (Ac-
cessed Jul 2019).
[Eckhouse et al., 2019] Eckhouse, L., Lum, K., Conti-Cook, C., and Ciccolini, J.
184
(2019). Layers of bias: A unified approach for understanding problems with
risk assessment. Criminal Justice and Behavior, 46(2):185–209.
[Edelman et al., 2017] Edelman, B., Luca, M., and Svirsky, D. (2017). Racial
discrimination in the sharing economy: Evidence from a field experiment.
American Economic Journal: Applied Economics, 9(2):1–22.
[Edelman and Luca, 2014] Edelman, B. G. and Luca, M. (2014). Digital discrimi-
nation: The case of airbnb. com. Harvard Business School NOM Unit Working
Paper, (14-054).
[Edmondson et al., 2004] Edmondson, A. C., Kramer, R. M., and Cook, K. S.
(2004). Psychological safety, trust, and learning in organizations: A group-level
lens. Trust and Distrust in Organizations: Dilemmas and Approaches, 12:239–272.
[Ellison and Hancock, 2013] Ellison, N. B. and Hancock, J. T. (2013). Profile as
promise: Honest and deceptive signals in online dating. IEEE Security and
Privacy, 11(5):84–88.
[Ellison et al., 2012] Ellison, N. B., Hancock, J. T., and Toma, C. L. (2012). Profile
as promise: A framework for conceptualizing veracity in online dating self-
presentations. New Media & Society, 14(1):45–62.
[Ellison et al., 2014] Ellison, N. B., Vitak, J., Gray, R., and Lampe, C. (2014). Culti-
vating social resources on social network sites: Facebook relationship mainte-
nance behaviors and their role in social capital processes. Journal of Computer-
Mediated Communication, 19(4):855–870.
[Ely, 1980] Ely, J. H. (1980). Democracy and Distrust: A Theory of Judicial Review.
Harvard University Press.
[Emerson, 1976] Emerson, R. M. (1976). Social exchange theory. Annual Review
of Sociology, pages 335–362.
[Epstein et al., 2019] Epstein, Z., Pennycook, G., and Rand, D. (2019). Letting the
crowd steer the algorithm: Laypeople can effectively identify misinformation
sources. PsyArXiv. Working Paper.
[Ert et al., 2016] Ert, E., Fleischer, A., and Magen, N. (2016). Trust and reputation
in the sharing economy: The role of personal photos in airbnb. Tourism
Management, 55:62–73.
185
[Eslami et al., 2015] Eslami, M., Rickman, A., Vaccaro, K., Aleyasen, A., Vuong,
A., Karahalios, K., Hamilton, K., and Sandvig, C. (2015). I always assumed that
i wasn’t really that close to [her]: Reasoning about invisible algorithms in news
feeds. In Proceedings of the ACM Conference on Human Factors in Computing
Systems, pages 153–162. ACM.
[Eslami et al., 2019] Eslami, M., Vaccaro, K., Gilbert, E., Lee, M. K., and Kara-
halios, K. (2019). User attitudes towards algorithmic opacity and transparency
in online reviewing platforms. In Proceedings of the ACM Conference on Human
Factors in Computing Systems. ACM.
[Facebook, 2018] Facebook (2018). Facebook help center. https://www.
facebook.com/help/1629740080681586. (Accessed Sep 2018).
[Ferguson and Peterson, 2015] Ferguson, A. J. and Peterson, R. S. (2015). Sinking
slowly: Diversity in propensity to trust predicts downward trust spirals in
small groups. Journal of Applied Psychology, 100(4):1012.
[Filippas et al., 2018] Filippas, A., Horton, J., and Golden, J. (2018). Reputation
inflation: Evidence from an online labor market. In Proceedings of the ACM
Conference on Economics and Computation.
[Fine and Holyfield, 1996] Fine, G. A. and Holyfield, L. (1996). Secrecy, trust,
and dangerous leisure: Generating group cohesion in voluntary organizations.
Social Psychology Quarterly, pages 22–38.
[Flintham et al., 2018] Flintham, M., Karner, C., Bachour, K., Creswick, H.,
Gupta, N., and Moran, S. (2018). Falling for fake news: investigating the
consumption of news via social media. In Proceedings of the ACM Conference on
Human Factors in Computing Systems, page 376. ACM.
[Fogel and Nehmad, 2009] Fogel, J. and Nehmad, E. (2009). Internet social net-
work communities: Risk taking, trust, and privacy concerns. Computers in
Human Behavior, 25(1):153–160.
[Forest and Wood, 2012] Forest, A. L. and Wood, J. V. (2012). When social net-
working is not working: Individuals with low self-esteem recognize but do
not reap the benefits of self-disclosure on facebook. Psychological Science,
23(3):295–302.
[Fradkin et al., 2015] Fradkin, A., Grewal, E., Holtz, D., and Pearson, M. (2015).
Bias and reciprocity in online reviews: Evidence from field experiments on
186
airbnb. In Proceedings of the ACM Conference on Economics and Computation,
pages 641–641. ACM.
[Fukuyama, 1995] Fukuyama, F. (1995). Trust: The Social Virtues and the Creation
of Prosperity. Free Press.
[Gambetta, 1988] Gambetta, D. (1988). Trust: Making and Breaking Cooperative
Relations. Basil Blackwell.
[Gambetta, 2009] Gambetta, D. (2009). Codes of the Underworld: How Criminals
Communicate. Princeton University Press.
[Gambetta and Hamill, 2005] Gambetta, D. and Hamill, H. (2005). Streetwise:
How Taxi Drivers Establish Customer’s Trustworthiness. Russell Sage Foundation.
[Gee, 2017] Gee, K. (2017). In unilever’s radical hiring experiment, resumes are
out, algorithms are in. Wall Street Journal, 26.
[Gefen and Straub, 2004] Gefen, D. and Straub, D. W. (2004). Consumer trust
in b2c e-commerce and the importance of social presence: experiments in
e-products and e-services. Omega, 32(6):407–424.
[Gibbs et al., 2010] Gibbs, J. L., Ellison, N. B., and Lai, C.-H. (2010). First comes
love, then comes google: An investigation of uncertainty reduction strate-
gies and self-disclosure in online dating. Communication Research, page
0093650210377091.
[Gilbert and Karahalios, 2009] Gilbert, E. and Karahalios, K. (2009). Predicting
tie strength with social media. In Proceedings of the ACM Conference on Human
Factors in Computing Systems, pages 211–220. ACM.
[Gill et al., 2005] Gill, H., Boies, K., Finegan, J. E., and McNally, J. (2005). An-
tecedents of trust: Establishing a boundary condition for the relation between
propensity to trust and intention to trust. Journal of Business and Psychology,
19(3):287–302.
[Glaeser et al., 2000] Glaeser, E. L., Laibson, D. I., Scheinkman, J. A., and Soutter,
C. L. (2000). Measuring trust. The Quarterly Journal of Economics, 115(3):811–846.
[Goodfellow et al., 2016] Goodfellow, I., Bengio, Y., and Courville, A. (2016).
Deep Learning. MIT Press. http://www.deeplearningbook.org.
187
[Goodfellow et al., 2014] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative
adversarial nets. In Advances in Neural Information Processing Systems, pages
2672–2680.
[Google, 2016] Google (2016). Google translate api documentation. https:
//cloud.google.com/translate/.
[Google, 2019] Google (2019). Google merchant center: Image guide-
lines. https://support.google.com/merchants/answer/6324350?
hl=en&ref_topic=6324338. (Accessed Jul 2019).
[Goswami et al., 2011] Goswami, A., Chittar, N., and Sung, C. H. (2011). A
study on the impact of product images on user clicks for online shopping.
In Companion to the International Conference on World Wide Web, pages 45–46.
ACM.
[Grabner-Kraeuter, 2002] Grabner-Kraeuter, S. (2002). The role of consumers’
trust in online-shopping. Journal of Business Ethics, 39(1-2):43–50.
[Granovetter, 1985] Granovetter, M. (1985). Economic action and social structure:
The problem of embeddedness. American Journal of Sociology, 91(3):481–510.
[Granovetter, 1973] Granovetter, M. S. (1973). The strength of weak ties. Ameri-
can Journal of Sociology, 78(6):1360–1380.
[Grbovic and Cheng, 2018] Grbovic, M. and Cheng, H. (2018). Real-time person-
alization using embeddings for search ranking at airbnb. In Proceedings of the
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
pages 311–320. ACM.
[Grinberg et al., 2019] Grinberg, N., Joseph, K., Friedland, L., Swire-Thompson,
B., and Lazer, D. (2019). Fake news on twitter during the 2016 us presidential
election. Science, 363(6425):374–378.
[Gubin et al., 2017] Gubin, M., Kao, W., Vickrey, D., and Maykov, A. (2017).
News feed ranking model based on social information of viewer. US Patent
9,582,786.
[Guillory and Hancock, 2012] Guillory, J. and Hancock, J. T. (2012). The effect
of linkedin on deception in resumes. Cyberpsychology, Behavior, and Social
Networking, 15(3):135–140.
188
[Gulati, 1995] Gulati, R. (1995). Does familiarity breed trust? the implications
of repeated ties for contractual choice in alliances. Academy of Management
Journal, 38(1):85–112.
[Gunther, 1988] Gunther, A. (1988). Attitude extremity and trust in media. Jour-
nalism Quarterly, 65(2):279–287.
[Guttentag, 2015] Guttentag, D. (2015). Airbnb: Disruptive innovation and the
rise of an informal tourism accommodation sector. Current Issues in Tourism,
18(12):1192–1217.
[Hamari et al., 2016] Hamari, J., Sjöklint, M., and Ukkonen, A. (2016). The shar-
ing economy: Why people participate in collaborative consumption. Journal of
the Association for Information Science and Technology, 67(9):2047–2059.
[Hannak et al., 2014] Hannak, A., Margolin, D., Keegan, B., and Weber, I. (2014).
Get back! you don’t know me like that: The social mediation of fact checking
interventions in twitter conversations. In Proceedings of the International AAAI
Conference on Weblogs and Social Media.
[Hannak et al., 2017] Hannak, A., Wagner, C., Garcia, D., Mislove, A.,
Strohmaier, M., and Wilson, C. (2017). Bias in online freelance marketplaces:
Evidence from taskrabbit and fiverr. In Proceedings of the ACM Conference
on Computer-Supported Cooperative Work & Social Computing, pages 1914–1933.
ACM.
[Hardin, 1999] Hardin, R. (1999). Do we want trust in government. Democracy
and Trust, pages 22–41.
[Hardin, 2002] Hardin, R. (2002). Trust and Trustworthiness. Russell Sage Foun-
dation.
[Hether et al., 2014] Hether, H. J., Murphy, S. T., and Valente, T. W. (2014). It’s
better to give than to receive: The role of social support, trust, and participation
on health-related social networking sites. Journal of Health Communication,
19(12):1424–1439.
[Hogg, 1993] Hogg, M. A. (1993). Group cohesiveness: A critical review and
some new directions. European Review of Social Psychology, 4(1):85–111.
[Hohenstein and Jung, 2018] Hohenstein, J. and Jung, M. (2018). Ai-supported
messaging: An investigation of human-human text conversation with ai sup-
189
port. In Companion to the ACM Conference on Human Factors in Computing
Systems, page LBW089. ACM.
[Holtz et al., 2017] Holtz, D., Lynn MacLean, D., and Aral, S. (2017). Social
structure and trust in massive digital markets.
[Homans, 1958] Homans, G. C. (1958). Social behavior as exchange. American
Journal of Sociology, pages 597–606.
[Hong and Cho, 2011] Hong, I. B. and Cho, H. (2011). The impact of consumer
trust on attitudinal loyalty and purchase intentions in b2c e-marketplaces: In-
termediary trust vs. seller trust. International Journal of Information Management,
31(5):469–479.
[Hutson et al., 2018] Hutson, J. A., Taft, J. G., Barocas, S., and Levy, K. (2018).
Debiasing desire: Addressing bias & discrimination on intimate platforms.
Proceedings of the ACM on Human-Computer Interaction, 2(CSCW):73.
[Ikkala and Lampinen, 2015] Ikkala, T. and Lampinen, A. (2015). Monetizing
network hospitality: Hospitality and sociability in the context of airbnb. In
Proceedings of the ACM Conference on Computer-Supported Cooperative Work &
Social Computing, pages 1033–1044. ACM.
[Inside Airbnb., 2016] Inside Airbnb. (2016). About inside airbnb. http://
insideairbnb.com/about.html. (Accessed May 2016).
[Jakesch et al., 2019] Jakesch, M., French, M., Ma, X., Hancock, J. T., and Naaman,
M. (2019). Ai-mediated communication: How the perception that profile text
was written by ai affects trustworthiness. In Proceedings of the ACM Conference
on Human Factors in Computing Systems, volume 51, page 22.
[Jarvenpaa and Leidner, 1999] Jarvenpaa, S. L. and Leidner, D. E. (1999). Com-
munication and trust in global virtual teams. Organization Science, 10(6):791–
815.
[Johnson and Mislin, 2011] Johnson, N. D. and Mislin, A. A. (2011). Trust games:
A meta-analysis. Journal of Economic Psychology, 32(5):865–889.
[Johnson-George and Swap, 1982] Johnson-George, C. and Swap, W. C. (1982).
Measurement of specific interpersonal trust: Construction and validation
of a scale to assess trust in a specific other. Journal of Personality and Social
Psychology, 43(6):1306.
190
[Jones and Leonard, 2008] Jones, K. and Leonard, L. N. (2008). Trust in
consumer-to-consumer electronic commerce. Information & Management,
45(2):88–95.
[Jøsang et al., 2007] Jøsang, A., Ismail, R., and Boyd, C. (2007). A survey of trust
and reputation systems for online service provision. Decision Support Systems,
43(2):618–644.
[Jourard and Lasakow, 1958] Jourard, S. M. and Lasakow, P. (1958). Some factors
in self-disclosure. The Journal of Abnormal and Social Psychology, 56(1):91.
[Kahneman, 2011] Kahneman, D. (2011). Thinking, Fast and Slow. Macmillan.
[Kannan et al., 2016] Kannan, A., Kurach, K., Ravi, S., Kaufmann, T., Tomkins,
A., Miklos, B., Corrado, G., Lukacs, L., Ganea, M., Young, P., et al. (2016).
Smart reply: Automated response suggestion for email. In Proceedings of the
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
pages 955–964. ACM.
[Kantsperger and Kunz, 2010] Kantsperger, R. and Kunz, W. H. (2010). Con-
sumer trust in service companies: a multiple mediating analysis. Managing
Service Quality: An International Journal, 20(1):4–25.
[Karras et al., 2018] Karras, T., Laine, S., and Aila, T. (2018). A style-based
generator architecture for generative adversarial networks. arXiv preprint
arXiv:1812.04948.
[Kiapour et al., 2015] Kiapour, M. H., Han, X., Lazebnik, S., Berg, A. C., and
Berg, T. L. (2015). Where to buy it: Matching street clothing photos in online
shops. In Proceedings of the IEEE International Conference on Computer Vision and
Pattern Recognition. IEEE.
[Kiesler et al., 1984] Kiesler, S., Siegel, J., and McGuire, T. W. (1984). Social
psychological aspects of computer-mediated communication. American Psy-
chologist, 39(10):1123.
[Kiousis, 2001] Kiousis, S. (2001). Public trust or mistrust? perceptions of media
credibility in the information age. Mass Communication & Society, 4(4):381–403.
[Kiyonari et al., 2006] Kiyonari, T., Yamagishi, T., Cook, K. S., and Cheshire,
C. (2006). Does trust beget trustworthiness? trust and trustworthiness in
191
two games and two cultures: A research note. Social Psychology Quarterly,
69(3):270–283.
[Kizilcec, 2016] Kizilcec, R. F. (2016). How much information?: Effects of trans-
parency on trust in an algorithmic interface. In Proceedings of the ACM Confer-
ence on Human Factors in Computing Systems, pages 2390–2395. ACM.
[Kohring and Matthes, 2007] Kohring, M. and Matthes, J. (2007). Trust in news
media: Development and validation of a multidimensional scale. Communica-
tion Research, 34(2):231–252.
[Kramer et al., 2014] Kramer, A. D., Guillory, J. E., and Hancock, J. T. (2014).
Experimental evidence of massive-scale emotional contagion through social
networks. volume 111, pages 8788–8790. National Academy of Sciences.
[Kramer, 1999] Kramer, R. M. (1999). Trust and distrust in organizations: Emerg-
ing perspectives, enduring questions. Annual Review of Psychology, 50(1):569–
598.
[Kraut and Fiore, 2014] Kraut, R. E. and Fiore, A. T. (2014). The role of founders
in building online groups. In Proceedings of the ACM Conference on Computer-
Supported Cooperative Work & Social Computing, pages 722–732. ACM.
[Kraut and Resnick, 2012] Kraut, R. E. and Resnick, P. (2012). Building Successful
Online Communities: Evidence-Based Social Design. MIT Press.
[Kricheli-Katz and Regev, 2016] Kricheli-Katz, T. and Regev, T. (2016). How
many cents on the dollar? women and men in product markets. Science
Advances, 2(2):e1500599.
[Kuwabara, 2015] Kuwabara, K. (2015). Do reputation systems undermine trust?
divergent effects of enforcement type on generalized trust and trustworthi-
ness1. American Journal of Sociology, 120(5):1390–1428.
[La Macchia et al., 2016] La Macchia, S. T., Louis, W. R., Hornsey, M. J., and
Leonardelli, G. J. (2016). In small we trust: Lay theories about small and large
groups. Personality and Social Psychology Bulletin, 42(10):1321–1334.
[Lai and Tan, 2018] Lai, V. and Tan, C. (2018). On human predictions with ex-
planations and predictions of machine learning models: A case study on
deception detection. In Proceedings of the Conference on Fairness, Accountability,
and Transparency.
192
[Lampe et al., 2006] Lampe, C., Ellison, N., and Steinfield, C. (2006). A face
(book) in the crowd: Social searching vs. social browsing. In Proceedings of the
ACM Conference on Computer-Supported Cooperative Work & Social Computing,
pages 167–170. ACM.
[Lampinen and Cheshire, 2016] Lampinen, A. and Cheshire, C. (2016). Hosting
via airbnb: Motivations and financial assurances in monetized network hos-
pitality. In Proceedings of the ACM Conference on Human Factors in Computing
Systems, pages 1669–1680. ACM.
[Lankton and McKnight, 2011] Lankton, N. K. and McKnight, D. H. (2011).
What does it mean to trust facebook?: examining technology and interper-
sonal trust beliefs. ACM SIGMIS Database: The DATABASE for Advances in
Information Systems, 42(2):32–54.
[Larzelere and Huston, 1980] Larzelere, R. E. and Huston, T. L. (1980). The
dyadic trust scale: Toward understanding interpersonal trust in close relation-
ships. Journal of Marriage and the Family, pages 595–604.
[Lavergne and Mullainathan, 2004] Lavergne, M. and Mullainathan, S. (2004).
Are emily and greg more employable than lakisha and jamal? a field ex-
periment on labor market discrimination. The American Economic Review,
94(4):991–1013.
[Lee et al., 2015a] Lee, D., Hyun, W., Ryu, J., Lee, W. J., Rhee, W., and Suh, B.
(2015a). An analysis of social features associated with room sales of airbnb. In
Proceedings of the ACM Conference on Computer-Supported Cooperative Work &
Social Computing. ACM.
[Lee et al., 2013] Lee, J. J., Knox, B., Baumann, J., Breazeal, C., and DeSteno, D.
(2013). Computationally modeling interpersonal trust. Frontiers in Psychology,
4:893.
[Lee et al., 2015b] Lee, M. K., Kusbit, D., Metsky, E., and Dabbish, L. (2015b).
Working with machines: The impact of algorithmic and data-driven man-
agement on human workers. In Proceedings of the ACM Conference on Human
Factors in Computing Systems, pages 1603–1612. ACM.
[Levy and Barocas, 2017] Levy, K. and Barocas, S. (2017). Designing against
discrimination in online markets. Berkeley Technology Law Journal.
[Li et al., 2010] Li, C., Loui, A. C., and Chen, T. (2010). Towards aesthetics: A
193
photo quality assessment and photo selection system. In Proceedings of the
ACM International Conference on Multimedia, pages 827–830. ACM.
[Linden et al., 2003] Linden, G., Smith, B., and York, J. (2003). Amazon. com rec-
ommendations: Item-to-item collaborative filtering. IEEE Internet Computing,
(1):76–80.
[Liu et al., 2017] Liu, M., Ding, W., Park, D. H., Fang, Y., Yan, R., and Hu, X.
(2017). Which used product is more sellable? a time-aware approach. Informa-
tion Retrieval Journal, 20(2):81–108.
[Liu et al., 2016] Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y.,
and Berg, A. C. (2016). Ssd: Single shot multibox detector. In Proceedings of the
European Conference on Computer Vision, pages 21–37. Springer.
[Lu et al., 2014] Lu, X., Lin, Z., Jin, H., Yang, J., and Wang, J. Z. (2014). Rapid:
Rating pictorial aesthetics using deep learning. In Proceedings of the ACM
International Conference on Multimedia, pages 457–466. ACM.
[Luhmann, 1979] Luhmann, N. (1979). Trust and Power: Two Works by Niklas
Luhmann. Chichester: John Wiley.
[Ma et al., 2019a] Ma, X., Cheng, J., Iyer, S., and Naaman, M. (2019a). When
do people trust their social groups? In Proceedings of the ACM Conference on
Human Factors in Computing Systems. ACM.
[Ma et al., 2016] Ma, X., Hancock, J., and Naaman, M. (2016). Anonymity, inti-
macy and self-disclosure in social media. In Proceedings of the ACM Conference
on Human Factors in Computing Systems, pages 3857–3869. ACM.
[Ma et al., 2017a] Ma, X., Hancock, J. T., Lim Mingjie, K., and Naaman, M.
(2017a). Self-disclosure and perceived trustworthiness of airbnb host pro-
files. In Proceedings of the ACM Conference on Computer-Supported Cooperative
Work & Social Computing. ACM.
[Ma et al., 2017b] Ma, X., Hancock, J. T., Mingjie, K. L., and Naaman, M. (2017b).
Self-disclosure and perceived trustworthiness of airbnb host profiles. In
Proceedings of the ACM Conference on Computer-Supported Cooperative Work
& Social Computing. ACM.
[Ma et al., 2019b] Ma, X., Mezghani, L., Wilber, K., Hong, H., Piramuthu, R.,
Naaman, M., and Belongie, S. (2019b). Understanding image quality and trust
194
in peer-to-peer marketplaces. In Proceedings of the IEEE Winter Conference on
Applications of Computer Vision, pages 511–520. IEEE.
[Ma et al., 2017c] Ma, X., Neeraj, T., and Naaman, M. (2017c). A computational
approach to perceived trustworthiness of airbnb host profiles. In Eleventh
International AAAI Conference on Web and Social Media. AAAI.
[Marsh and Dibben, 2005] Marsh, S. and Dibben, M. R. (2005). Trust, untrust,
distrust and mistrust–an exploration of the dark (er) side. In Proceedings of the
International Conference on Trust Management, pages 17–33. Springer.
[Marwick and Boyd, 2011] Marwick, A. E. and Boyd, D. (2011). I tweet hon-
estly, i tweet passionately: Twitter users, context collapse, and the imagined
audience. New Media & Society, 13(1):114–133.
[Mayer and Davis, 1999] Mayer, R. C. and Davis, J. H. (1999). The effect of
the performance appraisal system on trust for management: A field quasi-
experiment. Journal of Applied Psychology, 84(1):123.
[Mayer et al., 1995] Mayer, R. C., Davis, J. H., and Schoorman, F. D. (1995). An
integrative model of organizational trust. Academy of Management Review,
20(3):709–734.
[Mayer and Gavin, 2005] Mayer, R. C. and Gavin, M. B. (2005). Trust in manage-
ment and performance: who minds the shop while the employees watch the
boss? Academy of Management Journal, 48(5):874–888.
[McEvily et al., 2003] McEvily, B., Perrone, V., and Zaheer, A. (2003). Trust as an
organizing principle. Organization Science, 14(1):91–103.
[McPherson et al., 2001] McPherson, M., Smith-Lovin, L., and Cook, J. M. (2001).
Birds of a feather: Homophily in social networks. Annual Review of Sociology,
27(1):415–444.
[Merrill and Cheshire, 2017] Merrill, N. and Cheshire, C. (2017). Trust your
heart: Assessing cooperation and trust with biosignals in computer-mediated
interactions. In Proceedings of the ACM Conference on Computer-Supported Coop-
erative Work & Social Computingg, pages 2–12. ACM.
[Meyerson et al., 1996] Meyerson, D., Weick, K. E., and Kramer, R. M. (1996).
Swift trust and temporary groups. Trust in Organizations: Frontiers of Theory
and Research, 166:195.
195
[Michailidou et al., 2008] Michailidou, E., Harper, S., and Bechhofer, S. (2008).
Visual complexity and aesthetic perception of web pages. In Proceedings of
the ACM International Conference on Design of Communication, pages 215–224.
ACM.
[Miller, 1974] Miller, A. H. (1974). Political issues and trust in government:
1964–1970. American Political Science Review, 68(03):951–972.
[Miller and Mitamura, 2003] Miller, A. S. and Mitamura, T. (2003). Are surveys
on trust trustworthy? Social Psychology Quarterly, pages 62–70.
[Miller, 2018] Miller, T. (2018). Explanation in artificial intelligence: Insights
from the social sciences. Artificial Intelligence.
[Misztal, 2013] Misztal, B. (2013). Trust in Modern Societies: The Search for the Bases
of Social Order. John Wiley & Sons.
[Mitra and Gilbert, 2014] Mitra, T. and Gilbert, E. (2014). The language that gets
people to give: Phrases that predict success on kickstarter. In Proceedings of the
ACM Conference on Computer-Supported Cooperative Work & Social Computing,
pages 49–61. ACM.
[Moser et al., 2017] Moser, C., Resnick, P., and Schoenebeck, S. (2017). Commu-
nity commerce: Facilitating trust in mom-to-mom sale groups on facebook.
In Proceedings of the ACM Conference on Human Factors in Computing Systems,
pages 4344–4357. ACM.
[Mukherjee et al., 2014] Mukherjee, A. N., Bhattacharyya, S., and Bera, R. (2014).
Role of information technology in human resource management of sme: A
study on the use of applicant tracking system. IBMRD’s Journal of Management
& Research, 3(1):1–22.
[Munro et al., 2010] Munro, R., Bethard, S., Kuperman, V., Lai, V. T., Melnick, R.,
Potts, C., Schnoebelen, T., and Tily, H. (2010). Crowdsourcing and language
studies: the new generation of linguistic data. In Proceedings of the NAACL HLT
2010 Workshop on Creating Speech and Language Data with Amazon’s Mechanical
Turk, pages 122–130.
[Murray et al., 2012] Murray, N., Marchesotti, L., and Perronnin, F. (2012). AVA:
a large-scale database for aesthetic visual analysis. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 2408–2415. IEEE.
196
[Naaman et al., 2019] Naaman, M., Levy, K., and Hancock, J. (2019). Ai-
mediated communication: Definition, research agenda, and ethical considera-
tions. Working Paper.
[Nannestad, 2008] Nannestad, P. (2008). What have we learned about general-
ized trust, if anything? Annual Review of Political Science, 11:413–436.
[Netzer et al., 2016] Netzer, O., Lemaire, A., and Herzenstein, M. (2016). When
words sweat: Identifying signals for loan default in the text of loan applica-
tions. In Advances in Consumer Research. Association for Consumer Research.
[Newman and Antin, 2016] Newman, R. and Antin, J. (2016). Building for trust:
Insights from our efforts to distill the fuel for the sharing economy. http:
//nerds.airbnb.com/building-for-trust. (Accessed May 2016).
[Nyhan, 2000] Nyhan, R. C. (2000). Changing the paradigm: Trust and its role
in public sector organizations. The American Review of Public Administration,
30(1):87–109.
[Ott et al., 2011] Ott, M., Choi, Y., Cardie, C., and Hancock, J. T. (2011). Finding
deceptive opinion spam by any stretch of the imagination. In Proceedings of the
Annual Meeting of the Association for Computational Linguistics, pages 309–319.
Association for Computational Linguistics.
[Paszke et al., 2017] Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E.,
DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and Lerer, A. (2017). Automatic
differentiation in pytorch. In Proceedings of the Workshop on Neural Information
Processing Systems.
[Pavlou and Dimoka, 2006] Pavlou, P. A. and Dimoka, A. (2006). The nature
and role of feedback text comments in online marketplaces: Implications for
trust building, price premiums, and seller differentiation. Information Systems
Research, 17(4):392–414.
[Paxton, 2007] Paxton, P. (2007). Association memberships and generalized trust:
A multilevel model across 31 countries. Social Forces, 86(1):47–76.
[Pennebaker et al., 2001] Pennebaker, J. W., Francis, M. E., and Booth, R. J. (2001).
Linguistic inquiry and word count: Liwc 2001. Mahway: Lawrence Erlbaum
Associates, 71(2001):2001.
[Pennycook and Rand, 2019] Pennycook, G. and Rand, D. G. (2019). Fighting
197
misinformation on social media using crowdsourced judgments of news
source quality. Proceedings of the National Academy of Sciences, 116(7):2521–
2526.
[Perez, 2018] Perez, S. (2018). Facebook is launching a new groups
tab and plug-in. https://techcrunch.com/2018/05/01/
facebook-is-launching-a-new-groups-tab-and-plugin. (Ac-
cessed Sep 2018).
[Pew, 2007] Pew (2007). Americans and social trust: Who, where
and why. http://www.pewsocialtrends.org/2007/02/22/
americans-and-social-trust-who-where-and-why. (Accessed
Sep 2018).
[Pew, 2019] Pew (2019). Public trust in government: 1958-
2019. https://www.people-press.org/2019/04/11/
public-trust-in-government-1958-2019/. (Accessed Jul 2019).
[Popat et al., 2016] Popat, K., Mukherjee, S., Strötgen, J., and Weikum, G. (2016).
Credibility assessment of textual claims on the web. In Proceedings of the
ACM International Conference on Information and Knowledge Management, pages
2173–2178. ACM.
[Preece and Maloney-Krichmar, 2005] Preece, J. and Maloney-Krichmar, D.
(2005). Online communities: Design, theory, and practice. Journal of Computer-
Mediated Communication, 10(4).
[Putnam, 1993] Putnam, R. D. (1993). The prosperous community. The American
Prospect, 4(13):35–42.
[Putnam, 1995] Putnam, R. D. (1995). Bowling alone: America’s declining social
capital. Journal of Democracy, 6(1):65–78.
[Putnam, 2000] Putnam, R. D. (2000). Bowling Alone: The Collapse and Revival of
American Democracy. Simon and Schuster Nova York.
[Qiu et al., 2018] Qiu, W., Parigi, P., and Abrahao, B. (2018). More stars or more
reviews? In Proceedings of the ACM Conference on Human Factors in Computing
Systems, page 153. ACM.
[Rainie and Wellman, 2012] Rainie, L. and Wellman, B. (2012). Networked: The
New Social Operating System. Mit Press.
198
[Rashkin et al., 2017] Rashkin, H., Choi, E., Jang, J. Y., Volkova, S., and Choi, Y.
(2017). Truth of varying shades: Analyzing language in fake news and political
fact-checking. In Proceedings of the Conference on Empirical Methods in Natural
Language Processing, pages 2931–2937.
[Reinecke et al., 2013] Reinecke, K., Yeh, T., Miratrix, L., Mardiko, R., Zhao, Y.,
Liu, J., and Gajos, K. Z. (2013). Predicting users’ first impressions of website
aesthetics with a quantification of perceived visual complexity and colorful-
ness. In Proceedings of the ACM Conference on Human Factors in Computing
Systems, pages 2049–2058. ACM.
[Ren et al., 2007] Ren, Y., Kraut, R., and Kiesler, S. (2007). Applying common
identity and bond theory to design of online communities. Organization Studies,
28(3):377–408.
[Resnick et al., 2000] Resnick, P., Kuwabara, K., Zeckhauser, R., and Friedman,
E. (2000). Reputation systems. Communications of the ACM, 43(12):45–48.
[Resnick and Zeckhauser, 2002] Resnick, P. and Zeckhauser, R. (2002). Trust
among strangers in internet transactions: Empirical analysis of ebay’s repu-
tation system. In The Economics of the Internet and E-commerce, pages 127–157.
Emerald Group Publishing Limited.
[Resnick et al., 2006] Resnick, P., Zeckhauser, R., Swanson, J., and Lockwood, K.
(2006). The value of reputation on ebay: A controlled experiment. Experimental
Economics, 9(2):79–101.
[Ridings et al., 2002] Ridings, C. M., Gefen, D., and Arinze, B. (2002). Some
antecedents and effects of trust in virtual communities. The Journal of Strategic
Information Systems, 11(3-4):271–295.
[Rini, 2019] Rini, R. (2019). Deepfakes are coming. we can no longer be-
lieve what we see. https://www.nytimes.com/2019/06/10/opinion/
deepfake-pelosi-video.html. (Accessed Jul 2019).
[Ritzer, 1975] Ritzer, G. (1975). Sociology: A multiple paradigm science. The
American Sociologist, pages 156–167.
[Rosenblat, 2018] Rosenblat, A. (2018). Uberland: How Algorithms Are Rewriting
the Rules of Work. University of California Press.
[Rosenblat and Stark, 2016] Rosenblat, A. and Stark, L. (2016). Algorithmic labor
199
and information asymmetries: A case study of uber’s drivers. International
Journal of Communication.
[Rother et al., 2004] Rother, C., Kolmogorov, V., and Blake, A. (2004). Grabcut:
Interactive foreground extraction using iterated graph cuts. In Proceedings of
the ACM Transactions on Graphics, volume 23, pages 309–314. ACM.
[Rotter, 1967] Rotter, J. B. (1967). A new scale for the measurement of interper-
sonal trust. Journal of Personality, 35(4):651–665.
[Rotter, 1971] Rotter, J. B. (1971). Generalized expectancies for interpersonal
trust. American Psychologist, 26(5):443.
[Rousseau and Greller, 1994] Rousseau, D. M. and Greller, M. M. (1994). Hu-
man resource practices: Administrative contract makers. Human Resource
Management, 33(3):385–401.
[Rubin, 1975] Rubin, Z. (1975). Disclosing oneself to a stranger: Reciprocity and
its limits. Journal of Experimental Social Psychology, 11(3):233–260.
[Schoorman et al., 2007] Schoorman, F. D., Mayer, R. C., and Davis, J. H. (2007).
An integrative model of organizational trust: Past, present, and future.
Academy of Management Review.
[Schwarz et al., 2016] Schwarz, K., Wieschollek, P., and Lensch, H. (2016). Will
people like your image? learning the aesthetic space. In Proceedings of the IEEE
Winter Conference on Applications of Computer Vision.
[Shao et al., 2018] Shao, C., Ciampaglia, G. L., Varol, O., Yang, K.-C., Flammini,
A., and Menczer, F. (2018). The spread of low-credibility content by social bots.
Nature Communications, 9(1):4787.
[Shu et al., 2019a] Shu, K., Cui, L., Wang, S., Lee, D., and Liu, H. (2019a). De-
fend: Explainable fake news detection. In Proceedings of the ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining.
[Shu et al., 2017] Shu, K., Sliva, A., Wang, S., Tang, J., and Liu, H. (2017). Fake
news detection on social media: A data mining perspective. ACM SIGKDD
Explorations Newsletter, 19(1):22–36.
[Shu et al., 2019b] Shu, K., Wang, S., and Liu, H. (2019b). Beyond news contents:
200
The role of social context for fake news detection. In Proceedings of the ACM
International Conference on Web Search and Data Mining, pages 312–320. ACM.
[Shu et al., 2018] Shu, K., Zhou, X., Wang, S., Zafarani, R., and Liu, H. (2018).
Understanding user profiles on social media for fake news detection. In
Proceedings of the IEEE Conference on Multimedia Information Processing and
Retrieval.
[Silverman et al., 2018] Silverman, C., Lytvynenko, J., and Thuy Vo,
L. (2018). How facebook groups are being exploited to
spread misinformation, plan harassment, and radicalize people.
https://www.buzzfeednews.com/article/craigsilverman/
how-facebook-groups-are-being-exploited-to-spread. (Ac-
cessed Jul 2019).
[Simon and Brown, 1987] Simon, B. and Brown, R. (1987). Perceived intragroup
homogeneity in minority-majority contexts. Journal of Personality and Social
Psychology.
[Spence, 2002] Spence, M. (2002). Signaling in retrospect and the informational
structure of markets. The American Economic Review, 92(3):434–459.
[State et al., 2016] State, B., Abrahao, B. D., and Cook, K. (2016). Power imbal-
ance and rating systems. In Proceedings of the International AAAI Conference on
Web and Social Media, pages 368–377.
[Stokes, 1983] Stokes, J. P. (1983). Components of group cohesion: Intermember
attraction, instrumental value, and risk taking. Small Group Behavior.
[Sun, 2010] Sun, H. (2010). Sellers’ trust and continued use of online market-
places. Journal of the Association for Information Systems, 11(4):2.
[Sundar, 1998] Sundar, S. S. (1998). Effect of source attribution on perception of
online news stories. Journalism & Mass Communication Quarterly, 75(1):55–68.
[Susser et al., 2018] Susser, D., Roessler, B., and Nissenbaum, H. (2018). Online
manipulation: Hidden influences in a digital world. Working Paper.
[Szegedy et al., 2016] Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna,
Z. (2016). Rethinking the inception architecture for computer vision. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pages 2818–2826.
201
[Tan et al., 2014] Tan, C., Lee, L., and Pang, B. (2014). The effect of wording on
message propagation: Topic- and author-controlled natural experiments on
twitter. In Proceedings of the Annual Meeting of the Association for Computational
Linguistics.
[Tavakolifard and Almeroth, 2012] Tavakolifard, M. and Almeroth, K. C. (2012).
Social computing: an intersection of recommender systems, trust/reputation
systems, and social networks. IEEE Network, 26(4):53–58.
[Taylor et al., 2007] Taylor, P., Funk, C., and Clark, A. (2007). Americans and
social trust: Who, where and why. A Social Trends Report.
[Thebault-Spieker et al., 2017] Thebault-Spieker, J., Terveen, L., and Hecht, B.
(2017). Toward a geographic understanding of the sharing economy: Sys-
temic biases in uberx and taskrabbit. ACM Transactions on Computer-Human
Interaction, 24(3):21.
[Thebault-Spieker et al., 2015] Thebault-Spieker, J., Terveen, L. G., and Hecht, B.
(2015). Avoiding the south side and the suburbs: The geography of mobile
crowdsourcing markets. In Proceedings of the ACM Conference on Computer-
Supported Cooperative Work & Social Computing, CSCW ’15, pages 265–275.
ACM.
[Toma and Hancock, 2012] Toma, C. L. and Hancock, J. T. (2012). What lies
beneath: The linguistic traces of deception in online dating profiles. Journal of
Communication, 62(1):78–97.
[Toma et al., 2008] Toma, C. L., Hancock, J. T., and Ellison, N. B. (2008). Separat-
ing fact from fiction: An examination of deceptive self-presentation in online
dating profiles. Personality and Social Psychology Bulletin, 34(8):1023–1036.
[Tully et al., 2019] Tully, M., Vraga, E. K., and Bode, L. (2019). Designing and
testing news literacy messages for social media. Mass Communication and
Society, pages 1–25.
[Ugander et al., 2012] Ugander, J., Backstrom, L., Marlow, C., and Kleinberg, J.
(2012). Structural diversity in social contagion. Proceedings of the National
Academy of Sciences, pages 5962–5966.
[Uski and Lampinen, 2014] Uski, S. and Lampinen, A. (2014). Social norms and
self-presentation on social network sites: Profile work in action. New Media &
Society.
202
[Uzzi, 1996] Uzzi, B. (1996). The sources and consequences of embeddedness
for the economic performance of organizations: The network effect. American
Sociological Review, pages 674–698.
[Vaidya et al., 2019] Vaidya, T., Votipka, D., Mazurek, M. L., and Sherr, M. (2019).
Does being verified make you more credible? account verification’s effect on
tweet credibility. In Proceedings of the ACM Conference on Human Factors in
Computing Systems. ACM.
[Van Vugt and Hart, 2004] Van Vugt, M. and Hart, C. M. (2004). Social identity
as social glue: the origins of group loyalty. Journal of Personality and Social
Psychology, 86(4):585.
[Vasilescu et al., 2012] Vasilescu, B., Capiluppi, A., and Serebrenik, A. (2012).
Gender, representation and online participation: A quantitative study of
stackoverflow. In Proceedings of the International Conference on Social Informatics,
pages 332–338. IEEE.
[Vigoda-Gadot and Talmud, 2010] Vigoda-Gadot, E. and Talmud, I. (2010). Or-
ganizational politics and job outcomes: The moderating effect of trust and
social support. Journal of Applied Social Psychology, 40(11):2829–2861.
[Vosoughi et al., 2018] Vosoughi, S., Roy, D., and Aral, S. (2018). The spread of
true and false news online. Science, 359(6380):1146–1151.
[Walther and Bunz, 2005] Walther, J. B. and Bunz, U. (2005). The rules of virtual
groups: Trust, liking, and performance in computer-mediated communication.
Journal of Communication, 55(4):828–846.
[Wang, 2019] Wang, P. (2019). This person does not exist. http://
thispersondoesnotexist.com. (Accessed Jul 2019).
[Wang et al., 2016] Wang, X., Sun, Z., Zhang, W., Zhou, Y., and Jiang, Y.-G. (2016).
Matching user photos to online products with robust deep features. In Pro-
ceedings of the ACM International Conference on Multimedia Retrieval, pages 7–14.
ACM.
[Wilber et al., 2017] Wilber, M. J., Fang, C., Jin, H., Hertzmann, A., Collomosse, J.,
and Belongie, S. (2017). Bam! the behance artistic media dataset for recognition
beyond photography. In Proceedings of the IEEE International Conference on
Computer Vision.
203
[Wu et al., 2010] Wu, G., Hu, X., and Wu, Y. (2010). Effects of perceived interac-
tivity, perceived web assurance and disposition to trust on initial online trust.
Journal of Computer-Mediated Communication, 16(1):1–26.
[WVS, 2018] WVS (2018). World values survey database. http://www.
worldvaluessurvey.org/WVSContents.jsp. (Accessed Aug 2018).
[Yakovleva et al., 2010] Yakovleva, M., Reilly, R. R., and Werko, R. (2010). Why
do we trust? moving beyond individual to dyadic perceptions. Journal of
Applied Psychology, 95(1):79.
[Yamagishi and Cook, 1993] Yamagishi, T. and Cook, K. S. (1993). Generalized
exchange and social dilemmas. Social Psychology Quarterly, pages 235–248.
[Yamagishi et al., 2009] Yamagishi, T., Matsuda, M., Yoshikai, N., Takahashi, H.,
and Usui, Y. (2009). Solving the lemons problem with reputation. In eTrust:
Forming Relationships in the Online World.
[Yamagishi and Yamagishi, 1994] Yamagishi, T. and Yamagishi, M. (1994). Trust
and commitment in the united states and japan. Motivation and Emotion,
18(2):129–166.
[Yao et al., 2017a] Yao, W., Dai, Z., Huang, R., and Caverlee, J. (2017a). Online
deception detection refueled by real world data collection. In Proceedings of the
International Conference Recent Advances in Natural Language Processing.
[Yao et al., 2017b] Yao, Y., Viswanath, B., Cryan, J., Zheng, H., and Zhao, B. Y.
(2017b). Automated crowdturfing attacks and defenses in online review
systems. In Proceedings of the ACM Conference on Computer and Communications
Security.
[Yu and Grauman, 2014] Yu, A. and Grauman, K. (2014). Fine-grained visual
comparisons with local learning. In Proceedings of the Conference on Computer
Vision and Pattern Recognition.
[Yu and Grauman, 2017] Yu, A. and Grauman, K. (2017). Semantic jitter: Dense
supervision for visual comparisons via synthetic images. In Proceedings of the
International Conference on Computer Vision.
[Yuki et al., 2005] Yuki, M., Maddux, W. W., Brewer, M. B., and Takemura, K.
(2005). Cross-cultural differences in relationship-and group-based trust. Per-
sonality and Social Psychology Bulletin, 31(1):48–62.
204
[Zelmer, 2003] Zelmer, J. (2003). Linear public goods experiments: A meta-
analysis. Experimental Economics, 6(3):299–310.
[Zervas et al., 2015] Zervas, G., Proserpio, D., and Byers, J. (2015). A first look
at online reputation on airbnb, where every stay is above average. Working
Paper.
[Zhang et al., 2018] Zhang, A. X., Ranganathan, A., Metz, S. E., Appling, S., Se-
hat, C. M., Gilmore, N., Adams, N. B., Vincent, E., Lee, J., Robbins, M., et al.
(2018). A structured response to misinformation: Defining and annotating
credibility indicators in news articles. In Proceedings of the International Con-
ference on World Wide Web, pages 603–612. International World Wide Web
Conferences Steering Committee.
[Zhang et al., 2007] Zhang, J., Ackerman, M. S., and Adamic, L. (2007). Expertise
networks in online communities: structure and algorithms. In Proceedings of
the International Conference on World Wide Web, pages 221–230. ACM.
[Zhao et al., 2012] Zhao, L., Lu, Y., Wang, B., Chau, P. Y., and Zhang, L. (2012).
Cultivating the sense of belonging and motivating user participation in virtual
communities: A social capital perspective. International Journal of Information
Management, 32(6):574–588.
[Zheng et al., 2009] Zheng, X. S., Chakraborty, I., Lin, J. J.-W., and Rauschen-
berger, R. (2009). Correlating low-level image statistics with users-rapid
aesthetic and affective judgments of web pages. In Proceedings of the ACM
Conference on Human Factors in Computing Systems, pages 1–10. ACM.
[Zhou et al., 2008] Zhou, Y., Wilkinson, D., Schreiber, R., and Pan, R. (2008).
Large-scale parallel collaborative filtering for the netflix prize. In International
conference on algorithmic applications in management, pages 337–348. Springer.
205