Explaining Hypergraph Neural Networks: From Local Explanations to Global Concepts

TL;DR

This work tackles the explainability gap for hypergraph neural networks by introducing SHypX, a model-agnostic post-hoc explainer that yields local explanations as subhypergraphs and global explanations via unsupervised concept extraction. The local explainer optimizes a joint objective that balances faithfulness (via $D_{KL}$ between full and subhypergraph predictions) and concision (via $|G_{sub}|_1$) using differentiable Gumbel-Softmax sampling over potential node-hyperedge links. SHypX achieves superior fidelity and conciseness compared to baselines on four real and four synthetic hypergraphs, and enables a tunable faithfulness-concision tradeoff, while the global explainer provides class-level explanations through representative concepts. The introduction of novel synthetic hypergraph benchmarks and generalized fidelity metrics strengthens evaluation and offers practical insights for debugging and deploying hyperGNNs in structure-rich domains.

Abstract

Hypergraph neural networks are a class of powerful models that leverage the message passing paradigm to learn over hypergraphs, a generalization of graphs well-suited to describing relational data with higher-order interactions. However, such models are not naturally interpretable, and their explainability has received very limited attention. We introduce SHypX, the first model-agnostic post-hoc explainer for hypergraph neural networks that provides both local and global explanations. At the instance-level, it performs input attribution by discretely sampling explanation subhypergraphs optimized to be faithful and concise. At the model-level, it produces global explanation subhypergraphs using unsupervised concept extraction. Extensive experiments across four real-world and four novel, synthetic hypergraph datasets demonstrate that our method finds high-quality explanations which can target a user-specified balance between faithfulness and concision, improving over baselines by 25 percent points in fidelity on average.

Figures (10)

• Figure 1: Visualization of our hypergraph explainer providing local and global explanations. (Top) Instance-level explanations are obtained by optimizing the subhypergraph structure using a loss function that incentivizes faithfulness (the explanation is able to reproduce the original prediction well) and concision (the explanation is as minimal as possible). (Bottom) Model-level explanations are obtained by combining the instance-level explainer with unsupervised concept extraction. After clustering the latent space into concepts, the closest node to each concept's center is picked as a representative and explained using the instance-level approach to produce concept and class-level explanations.
• Figure 2: (a)-(b) Illustrative fragments of the "base" component of our synthetic hypergraphs. They come in two flavours: random, and tree (which is deterministic). (c)-(e) Synthetic hypergraph motifs of the house, cycle, and grid varieties. The node colors indicate class labels, which are each distinct from the class assigned to base nodes. The anchor node, whereby each motif is attached to the base, is denoted with a black outline. (e) A small example hypergraph of the H-RandHouse family (pink edges denotes perturbations, gray denotes base hypergraph and yellow denotes attached motifs).
• Figure 3: Analysing the trade off between faithfulness and concision in various hypergraph explainers. The figure shows $\textrm{Fid}_{-}^\textrm{KL}$ vs. mean explanation size for two select hypergraphs on two datasets. While all the baselines obtains very little improvement in fidelity as we increase the explanation size, our model consistently obtains more faithful explanations at every size budget.
• Figure 4: Global concepts on H-RandHouse dataset. Class 0 is the base hypergraph, Class 1 is top-of-the-house, Class 2 is middle-of-the-house, and Class 3 is bottom-of-the-house. Concepts were extracted with 10 clusters, which sufficed to score well on the concept completeness metric (Appendix \ref{['sec:appendix_concept_extraction']}).
• Figure 5: Concepts for H-CommHouse.