EdgeMask-HGNN: Learning to Sparsify Hypergraphs in Hypergraph Neural Networks

TMLR Paper8114 Authors

Abstract:

Hypergraph Neural Networks (HGNNs) have achieved remarkable performance in various learning tasks involving hypergraphs. However, existing HGNNs consider the input hypergraph as fixed during training, ignoring the fact that real-world hypergraphs may often contain noisy hyperedges or links that are irrelevant for the downstream task. This makes them prone to overfitting, poor generalizability, and degrades their effectiveness on heterophilic hypergraphs. To address these issues, we propose EdgeMask-HGNN, a supervised sparsification method that learns a discrete subhypergraph under an explicit budget constraint. EdgeMask-HGNN offers two distinct sparsification schemes: a fine-grained sparsification and a coarse-grained sparsification, both trained end-to-end using supervision from the downstream task. Extensive experiments on node classification benchmarks demonstrate that EdgeMask-HGNN is effective on heterophilic hypergraphs. On more homophilic datasets, its performance is often comparable to strong baselines. Beyond node
classification, EdgeMask-HGNN also shows superior link prediction performance on existing link prediction benchmarks compared to full training and unsupervised sparsification baselines.

Submission Type: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=rmFry8VwFw

Changes Since Last Submission:

**Changes since last submission:**
Revision changes colored in blue.

Assigned Action Editor: ~Christopher_Morris1

Submission Number: 8114