Scholarlink

<p>Graph Neural Networks (GNNs) excel in structured data analysis but struggle with real-world heterogeneous graphs-characterized by distinct substructures and sparse inter-substructure connections-which hinders local-global information fusion, while existing methods, limited by the inherent locality of neighborhood aggregation or prohibitive computational costs, face critical challenges in efficient and robust global modeling of structurally heterogeneous graphs; this paper proposes <i>MPA-GNN</i> with two core designs: multi-prior anchor initialization to cover heterogeneous substructures and a sparse Node<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\rightleftarrows \)</EquationSource> </InlineEquation>Anchor mechanism for low-cost inter-substructure propagation, and experiments on heterogeneous datasets show its competitive or superior performance, while ablation studies confirm the contributions of key modules, making <i>MPA-GNN</i> a unified, efficient solution for robust heterogeneous graph representation learning. Source code of MPA-GNN is freely available at: <a href="https://github.com/wuyuhang1107-arch/MPA-GNN">https://github.com/wuyuhang1107-arch/MPA-GNN</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

Multi-prior anchored graph neural networks for robust and adaptive representation learning

  • 期刊论文

  • 发表日期：

    2026年7月31日

查看全文
  • Bojun Xie,
  • Yuhang Wu,
  • Junfen Chen

摘要

Graph Neural Networks (GNNs) excel in structured data analysis but struggle with real-world heterogeneous graphs-characterized by distinct substructures and sparse inter-substructure connections-which hinders local-global information fusion, while existing methods, limited by the inherent locality of neighborhood aggregation or prohibitive computational costs, face critical challenges in efficient and robust global modeling of structurally heterogeneous graphs; this paper proposes MPA-GNN with two core designs: multi-prior anchor initialization to cover heterogeneous substructures and a sparse Node \(\rightleftarrows \) Anchor mechanism for low-cost inter-substructure propagation, and experiments on heterogeneous datasets show its competitive or superior performance, while ablation studies confirm the contributions of key modules, making MPA-GNN a unified, efficient solution for robust heterogeneous graph representation learning. Source code of MPA-GNN is freely available at: https://github.com/wuyuhang1107-arch/MPA-GNN.