Interpretable Hypergraph Neural Additive Networks

ICLR 2026 Conference Submission15274 Authors

Keywords: Hypergraph Learning, Interpretability, Generalized Additive Models

Abstract:

Hypergraph neural networks have emerged as a powerful framework for learning from higher-order structured data, where relationships among entities extend beyond pairwise connections. However, most current hypergraph neural networks are black-boxes that rely on post-hoc explanation methods to provide model insights. Such post-hoc explanations can be unreliable in high-stakes scenarios and knowledge discovery tasks. We introduce an inherently interpretable hypergraph neural additive network (HGNAN), an
extension of generalized additive models that facilitates interpretability in complex, higher-order relational learning settings. HGNAN provides clear visualizations of both global and local behaviors at the node and hyperedge levels while preserving the expressive power of hypergraphs. We evaluate HGNAN on node classification and hyperedge prediction across various datasets, achieving competitive performance compared to state-of-the-art methods. HGNAN also significantly outperforms existing approaches in recovering missing reactions in metabolic networks, while offering interpretablebiological insights into metabolic processes.

Primary Area: interpretability and explainable AI

Submission Number: 15274