<p>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%<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10462_2025_11199_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(-\)</EquationSource> </InlineEquation>32.37% and 2.23%<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10462_2025_11199_Article_IEq2.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(-\)</EquationSource> </InlineEquation>27.54% on hypergraph exploration and optimization tasks, and 1.19%<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10462_2025_11199_Article_IEq3.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(-\)</EquationSource> </InlineEquation>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: <a href="https://github.com/MingyuanLi88888/SHSL">https://github.com/MingyuanLi88888/SHSL.</a></p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Self-supervised hypergraph structure learning

  • Mingyuan Li,
  • Yanlin Yang,
  • Lei Meng,
  • Lu Peng,
  • Haixing Zhao,
  • Zhonglin Ye

摘要

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.