Learning to Explain Hypergraph Neural Networks

Abstract

Hypergraphs are expressive structures for describing higher-order relationships among entities, with widespread applications across biology and drug discovery. Hypergraph neural networks (HGNNs) have recently emerged as a promising representation learning approach on these structures for clustering, classification, and more. However, despite their promising performance, HGNNs remain a black box, and explaining how they make predictions remains an open challenge. To address this problem, we propose HyperEX, a post-hoc explainability framework for hypergraphs that can be applied to any trained HGNN. HyperEX computes node-hyperedge pair importance to identify sub-hypergraphs as explanations. Our experiments demonstrate how HyperEX learns important sub-hypergraphs responsible for driving node classification to give useful insight into HGNNs.