Graph Neural Networks Meet Probabilistic Graphical Models: A Survey

Abstract

Graphs are a powerful data structure for representing relational data, and Graph Neural Networks (GNNs) have emerged as effective tools for inference and learning on graph-structured data. Probabilistic Graphical Models (PGMs), which provide compact graphical representations of variable distributions, offer a complementary approach with well-developed methods for capturing relationships and conducting message passing. In this survey, we explore how PGMs can enhance GNNs. We discuss how GNNs benefit from structured representations in PGMs, generate explainable predictions, and infer relationships. We also examine how GNNs are used within PGMs for more efficient inference and structure learning.