Scholarlink

<p>Graph Neural Networks (GNNs) have achieved great success in learning node representations from graph-structured data. Most GNNs adopt a message-passing paradigm, where nodes act as information carriers and edges serve as communication channels. However, existing models typically represent edges using scalar weights, which only modulate the overall strength of messages and fail to distinguish the importance of individual feature dimensions. This coarse-grained formulation limits the expressiveness of information propagation. To address this issue, we propose to represent edges as vectors, enabling fine-grained, dimension-specific control over the message-passing process. We refer to these edge vectors as <i>relation representations</i>, as they capture the semantic relations between connected nodes. Building on this idea, we present a novel model, <i>Relation-Enhanced Feature Preference Network</i> (ReFPN), which integrates relation representation learning into the message-passing framework. ReFPN dynamically adjusts the contribution of each feature dimension based on the learned relation representations, thereby enhancing the effectiveness of information aggregation. Extensive experiments on five benchmark datasets demonstrate that ReFPN&#xa0;consistently outperforms strong baselines in both node classification and clustering tasks, validating the superiority of fine-grained relation-aware propagation.</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

Redefining edge representations for enhanced information propagation on GNNs

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Shengda Zhuo,
  • Tianbo Wang,
  • Lichun Li,
  • Zifeng Zhou,
  • Zelin Guan,
  • Yin Tang,
  • Min Chen,
  • Shuqiang Huang

摘要

Graph Neural Networks (GNNs) have achieved great success in learning node representations from graph-structured data. Most GNNs adopt a message-passing paradigm, where nodes act as information carriers and edges serve as communication channels. However, existing models typically represent edges using scalar weights, which only modulate the overall strength of messages and fail to distinguish the importance of individual feature dimensions. This coarse-grained formulation limits the expressiveness of information propagation. To address this issue, we propose to represent edges as vectors, enabling fine-grained, dimension-specific control over the message-passing process. We refer to these edge vectors as relation representations, as they capture the semantic relations between connected nodes. Building on this idea, we present a novel model, Relation-Enhanced Feature Preference Network (ReFPN), which integrates relation representation learning into the message-passing framework. ReFPN dynamically adjusts the contribution of each feature dimension based on the learned relation representations, thereby enhancing the effectiveness of information aggregation. Extensive experiments on five benchmark datasets demonstrate that ReFPN consistently outperforms strong baselines in both node classification and clustering tasks, validating the superiority of fine-grained relation-aware propagation.