Self-supervised hypergraph structure learning

Abstract

Traditional Hypergraph Neural Networks (HGNNs) often assume that hypergraph structures are perfectly constructed, yet real-world hypergraphs are typically corrupted by noise, missing data, or irrelevant information, limiting the effectiveness of hypergraph learning. To address this challenge, we propose SHSL, a novel Self-supervised Hypergraph Structure Learning framework that jointly explores and optimizes hypergraph structures without external labels. SHSL consists of two key components: a self-organizing initialization module that constructs latent hypergraph representations, and a differentiable optimization module that refines hypergraphs through gradient-based learning. These modules collaboratively capture high-order dependencies to enhance hypergraph representations. Furthermore, SHSL introduces a dual learning mechanism to simultaneously guide structure exploration and optimization within a unified framework. Experiments on six public datasets demonstrate that SHSL outperforms state-of-the-art baselines, achieving Accuracy improvements of 1.36%\(-\)32.37% and 2.23%\(-\)27.54% on hypergraph exploration and optimization tasks, and 1.19%\(-\)8.4% on non-hypergraph datasets. Robustness evaluations further validate SHSL’s effectiveness under noisy and incomplete scenarios, highlighting its practical applicability. The implementation of SHSL and all experimental codes are publicly available at: https://github.com/MingyuanLi88888/SHSL.