Adaptive Hypergraph Pruning with Learned Threshold Control and Attention-Based Contrastive Mining

TMLR Paper7346 Authors

Abstract: Hypergraph neural networks (HGNNs) effectively model multi-way interactions but suffer from severe scalability limitations due to quadratic computational costs across multiple behavioral contexts. Existing pruning approaches reduce computation using fixed, hand-crafted heuristics, which fail to adapt to diverse graph structures and often introduce representation distortions by removing semantically related nodes or creating spurious similarities that degrade contrastive learning. We propose \textbf{TriPrune-HGNN}, an adaptive hypergraph pruning framework with learnable mechanisms that eliminates manual hyperparameter tuning (over $80\%$ reduction) while achieving a superior accuracy--efficiency tradeoff. TriPrune-HGNN learns pruning schedules from graph statistics and training dynamics, adaptively mines informative contrastive pairs, and automatically balances competing learning objectives via meta-optimization. Extensive experiments on five benchmarks show that TriPrune-HGNN achieves the best overall accuracy--efficiency tradeoff among all evaluated methods, matching or exceeding competitive baselines on all predictive metrics, while reducing inference time by 72.3\% and memory usage by 81.1\% compared to unpruned models. Compared with efficient baselines of similar memory footprint, TriPrune-HGNN attains up to 5.6\% lower error. We note that methods optimized purely for speed (e.g., LightHGNN) achieve lower raw inference time at the cost of higher prediction error; our contribution is a favorable \emph{tradeoff} rather than dominance on every individual dimension.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Christopher_Morris1

Submission Number: 7346