HyperGC: Learning Hypergraph Representations via Full Hyperedge Reconstruction and Contrastive Evaluation
TL;DR
The proposed HyperGC is a hybrid self-supervised HRL framework that unifies generative and contrastive learning through two core strategies: (S1) full hyperedge reconstruction via progressive selection with dynamic target cross-entropy, and (S2) multi-view contrastive evaluation without handcrafted negative samples.
Abstract
Hypergraph representation learning (HRL) is essential for modeling complex groupwise relationships in real-world data. However, existing generative self-supervised HRL methods suffer from two key limitations: (C1) partial hyperedge reconstruction fails to capture high-order formation semantics, and (C2) negative-sampling-based evaluation is unstable and sensitive to sample quality. To address these issues, we propose HyperGC, a hybrid self-supervised HRL framework that unifies generative and contrastive learning through two core strategies: (S1) full hyperedge reconstruction via progressive selection with dynamic target cross-entropy, and (S2) multi-view contrastive evaluation without handcrafted negative samples. Experiments on 11 real-world hypergraph datasets across node classification, hyperedge prediction, and community detection show that each strategy is effective individually, while their combination consistently shows the best performance, outperforming 19 state-of-the-art HRL methods by up to 12.27%, 5.12%, and 34.35% on the respective tasks.