Generalized Reasoning with Graph Neural Networks by Relational Bayesian Network Encodings

Graph neural networks (GNNs) and statistical relational learning are two different approaches to learning with graph data. The former can provide highly accurate models for specific tasks when sufficient training data is available, whereas the latter supports a wider range of reasoning types, and can incorporate manual specifications of interpretable domain knowledge. In this paper we present a method to embed GNNs in a statistical relational learning framework, such that the predictive model represented by the GNN becomes part of a full generative model. This model then supports a wide range of queries, including general conditional probability queries, and computing most probable configurations of unobserved node attributes or edges. In particular, we demonstrate how this latter type of queries can be used to obtain model-level explanations of a GNN in a flexible and interactive manner.