Scholarlink

<p>The semi-supervised graph alignment problem aims to find the node correspondence across different graphs given a set of anchor links. Most existing methods employ the notion of alignment consistency or embedding-based techniques but overlook the global structure of graph data. Recently, an Optimal Transport (OT)-based method has been proposed for semi-supervised graph alignment by integrating structure-based embedding and OT distance, demonstrating its effectiveness in problem modeling. However, graphs to be aligned often exhibit significant structural differences, and a non-learnable transport cost design struggles to maintain generality when faced with such variations, especially in noisy real-world scenarios. Meanwhile, the challenge of efficiently incorporating anchor links into the cost design has not been thoroughly explored. In this paper, we propose <Emphasis FontCategory="NonProportional">RESAlign</Emphasis>, a <Emphasis Type="Underline">r</Emphasis>obust s<Emphasis Type="Underline">e</Emphasis>mi-<Emphasis Type="Underline">s</Emphasis>upervised graph <Emphasis Type="Underline">align</Emphasis>ment framework that addresses the cross-domain alignment problem from both direct and indirect perspectives. By integrating multiple objective functions and an anchor-assisted heterogeneous graph learning module into the design of the transport cost, our framework adapts to structural differences across various graphs. Moreover, an additional weight-sharing mechanism is introduced to address node alignment from a distinct perspective, enabling effective generalization to unsupervised scenarios. Finally, compared to eleven representative methods, the proposed model not only achieves outstanding performance but also demonstrates excellent robustness and efficiency.</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

Enhancing robust semi-supervised graph alignment via adaptive optimal transport

  • 期刊论文

  • 发表日期：

    2025年2月26日

查看全文
  • Songyang Chen,
  • Youfang Lin,
  • Yu Liu,
  • Yuwei Ouyang,
  • Zongshen Guo,
  • Lei Zou

摘要

The semi-supervised graph alignment problem aims to find the node correspondence across different graphs given a set of anchor links. Most existing methods employ the notion of alignment consistency or embedding-based techniques but overlook the global structure of graph data. Recently, an Optimal Transport (OT)-based method has been proposed for semi-supervised graph alignment by integrating structure-based embedding and OT distance, demonstrating its effectiveness in problem modeling. However, graphs to be aligned often exhibit significant structural differences, and a non-learnable transport cost design struggles to maintain generality when faced with such variations, especially in noisy real-world scenarios. Meanwhile, the challenge of efficiently incorporating anchor links into the cost design has not been thoroughly explored. In this paper, we propose RESAlign, a robust semi-supervised graph alignment framework that addresses the cross-domain alignment problem from both direct and indirect perspectives. By integrating multiple objective functions and an anchor-assisted heterogeneous graph learning module into the design of the transport cost, our framework adapts to structural differences across various graphs. Moreover, an additional weight-sharing mechanism is introduced to address node alignment from a distinct perspective, enabling effective generalization to unsupervised scenarios. Finally, compared to eleven representative methods, the proposed model not only achieves outstanding performance but also demonstrates excellent robustness and efficiency.