Scholarlink

<p>Heterophily in graph neural networks has received significant attention due to its ability to adapt to challenging graph-structured data. Most existing research focuses on supervised learning strategies, leveraging edge heterophily as a supervisory signal to guide message passing. However, these strategies are challenging to apply in more common unsupervised scenarios, where representing heterophilic graphs becomes even more difficult. Specifically, two challenges arise: capturing edge heterophily in graph-structured data under unsupervised conditions, and utilizing this information for effective representation learning. To address the first challenge, we introduce a new strategy called the Feature-Distribution Embedding based Edge-Heterophily Identification, which identifies edge heterophily by evaluating the similarity between feature-distribution embedding. Additionally, we propose an effective adaptive mechanism to enhance the capture of edge heterophily information. For the second challenge, we propose a latent-space cross-perspective contrastive strategy that improves the unsupervised optimization of the heterophily graph neural network, further refining the edge-heterophily identifications. To integrate these strategies, we propose a novel framework: <Emphasis Type="BoldUnderline">C</Emphasis>ross-perspective contrastive <Emphasis Type="BoldUnderline">U</Emphasis>nsupervised <Emphasis Type="BoldUnderline">G</Emphasis>raph <Emphasis Type="BoldUnderline">N</Emphasis>eural <Emphasis Type="BoldUnderline">N</Emphasis>etwork with edge heterophily discriminability (<b>CuGNN</b>). CuGNN effectively captures the edge heterophily by leveraging the feature-distribution embedding similarity of nodes in the absence of labels and uses this information to optimize node representations. This equips the model to handle complex heterophily graph structure in an unsupervised setting. Extensive experiments on 12 benchmark datasets across multiple learning scenarios demonstrate the superiority and robustness of CuGNN, highlighting its effectiveness in unsupervised heterophilic graph representation.</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

CuGNN: Cross-perspective contrastive learning for enhancing unsupervised GNNs beyond homophily

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Jiahui Yang,
  • Yi Wang,
  • Changqin Huang,
  • Qionghao Huang,
  • Xiaodi Huang

摘要

Heterophily in graph neural networks has received significant attention due to its ability to adapt to challenging graph-structured data. Most existing research focuses on supervised learning strategies, leveraging edge heterophily as a supervisory signal to guide message passing. However, these strategies are challenging to apply in more common unsupervised scenarios, where representing heterophilic graphs becomes even more difficult. Specifically, two challenges arise: capturing edge heterophily in graph-structured data under unsupervised conditions, and utilizing this information for effective representation learning. To address the first challenge, we introduce a new strategy called the Feature-Distribution Embedding based Edge-Heterophily Identification, which identifies edge heterophily by evaluating the similarity between feature-distribution embedding. Additionally, we propose an effective adaptive mechanism to enhance the capture of edge heterophily information. For the second challenge, we propose a latent-space cross-perspective contrastive strategy that improves the unsupervised optimization of the heterophily graph neural network, further refining the edge-heterophily identifications. To integrate these strategies, we propose a novel framework: Cross-perspective contrastive Unsupervised Graph Neural Network with edge heterophily discriminability (CuGNN). CuGNN effectively captures the edge heterophily by leveraging the feature-distribution embedding similarity of nodes in the absence of labels and uses this information to optimize node representations. This equips the model to handle complex heterophily graph structure in an unsupervised setting. Extensive experiments on 12 benchmark datasets across multiple learning scenarios demonstrate the superiority and robustness of CuGNN, highlighting its effectiveness in unsupervised heterophilic graph representation.