Answering Probabilistic Graph Queries from a single model

A Probabilistic Graph Query (PGQ) specifies target components of a graph to be predicted, given observed components as evidence. For example, a single-link query specifies a target link, with other links and node attributes as evidence. Previous works have developed customized models for different PGQ types; this work describes a unified framework for answering various types of PGQs based on an inductively trained probabilistic Generative Graph Model (GGM). Given a trained GGM and any user PGQ, our approach constructs, without retraining, a conditional model that outputs the PGQ probability. In this thesis, we utilize Variational Graph Auto-Encoders (VGAEs) as GGM with Conditional Variational Auto-Encoders (CVAEs) as conditional models. For evaluation, we apply query answering to inductive link prediction queries on six benchmark datasets. We find that for most datasets and query types, the VGAE predictions are better than baseline methods customized for link prediction.