Graphs are a natural representation for systems with interacting components (e.g. an online social network of users; a transaction network of bank accounts; an interaction network of proteins). As such, algorithms that predict node labels have wide-ranged applications from online content recommendation, fraud detection, to drug discovery. The traditional machine learning setting assumes data points are independently sampled, and thus makes predictions only based on each individual’s attributes. For interconnected vertices in an attributed graph, the correlation along the edges provide an additional source of information. To better understand and leverage those two types of information, we propose data models for attributed graphs that: (1) explain existing graph learning algorithms such as label propagation and graph convolutional network, (2) inspire new algorithms that achieves the state-of-the-art performances, (3) generate synthetic graph attributes that preserves characteristics in real-world data.