Deep generative models for subgraph prediction

Graph Neural Networks (GNNs) are crucial across various domains due to their ability to model complex relational data. This work introduces subgraph queries as a new task in deep graph learning. Unlike traditional tasks like link prediction or node classification, subgraph queries predict components of a target subgraph based on evidence from an observed subgraph. For instance, they can predict a set of target links and node labels. In this work, I have introduced VGAE+ to answer these queries, using a probabilistic deep Graph Generative Model (GGM): an inductively trained Variational Graph Auto-Encoder (VGAE), enhanced to represent a joint distribution over links, node features, and labels. Bayesian optimization tunes the weighting of links, node features, and labels. I developed deterministic and sampling-based inference methods for estimating subgraph probabilities from the VGAE distribution. Evaluation on six benchmark datasets shows VGAE+ surpasses baselines, with AUC improvements up to 0.2 points.