Joint Use of Node Attributes and Proximity for Semi-Supervised Classification on Graphs

The node classification problem is to infer unknown node labels in a graph given its structure and node attributes along with labels for some of the nodes. Approaches for this task typically assume that adjacent nodes have similar attributes and thus, that a node's label can be predicted from the labels of its neighbors. While such homophily is often observed (e.g., for political affiliation in social networks), the assumption may not hold for arbitrary graph datasets and classification tasks. In fact, nodes that share the same label may be adjacent but differ in their attributes; or may not be adjacent but have similar attributes. We aim to develop a node classification approach that can flexibly adapt to a range of settings wherein labels are correlated with graph structure, or node attributes, or both. To this end, we propose JANE (Jointly using Attributes and Node Embeddings): a novel and principled approach based on a generative probabilistic model that weighs the role of node proximity and attribute similarity in predicting labels. Our experiments on a variety of graph datasets and comparison with standard baselines demonstrate that JANE exhibits a superior combination of versatility and competitive performance.