People's daily actions and decisions are increasingly shaped by recommendation systems (recommenders) that selectively suggest and present information items, from e-commerce and content platforms to education and wellness applications. However, existing systems are often optimized to promote commercial metrics, such as click-through rates and sales, while overlooking utility for individual users. As a result, recommendations can be narrow, skewed, homogeneous, and divergent from users' aspirations. This thesis introduces \textbf{user-centric recommendation systems} that are built to optimize for individuals' benefit. These systems advance the state of the art of recommenders by addressing the bias and incompleteness of implicit feedback upon which existing systems rely, such as click, download, and share. Specifically, this thesis explores three research directions: (1) \textbf{Debiasing implicit feedback.} We leverage a Self-Normalized Inverse-Propensity-Scoring (SNIPS) technique to derive a debiased measure of recommendation performance. Our approach models and alleviates popularity bias and is shown to significantly reduce the Mean Absolute Error (MAE) of evaluating recommendation systems offline. (2) \textbf{Leveraging richer data sources to learn broader user preferences.} We develop an unsupervised learning algorithm to learn discriminative user representations from unstructured software usage traces. The learned representations significantly improve the performance of personalization systems for creative professionals, including creative content recommenders and user tagging systems. (3) \textbf{Interactive preference learning addressing the limitations of passively collected offline data.} We build an interactive learning framework to learn users' food preferences from adaptive pairwise comparisons. This framework enables a recipe recommender that satisfies users' tastes and nutritional expectations. We also design an onboarding survey to empower an intention-informed podcast recommender. Through lab and field experiments, we demonstrate that these systems can promote healthier diets and aspiration-aligned content choices. In addition to the aforementioned user-centric recommenders, this thesis also contributes an open source tool, named OpenRec, to tackle the challenges of model generalization and adaptation that arise in building heterogeneous recommendation systems. OpenRec provides modular interfaces so that monolithic algorithms can be readily decomposed or combined for diverse application scenarios. At the end of this thesis, we discuss future research to personalize pervasive intelligent systems for people and our society and to understand and mitigate the unintended consequences of personalization.