Scholarlink

<p>Heterogeneous Graph Neural Networks (HGNNs) excel in learning representations of Heterogeneous Graphs (HG) with rich structural and semantic information. However, many existing HGNN models tend to focus more on advantaged groups with abundant labeled data, while overlooking disadvantaged groups with fewer labels. This imbalance leads to the feature propagation process being dominated by the advantaged groups, which causes the disadvantaged groups to lose their unique characteristics during neighborhood aggregation. To address this fairness issue, termed label-aware fairness, we propose a novel framework, FairHGNN (Label-aware Fair Heterogeneous Graph Neural Network). FairHGNN leverages knowledge distillation to transfer neighborhood information from advantaged to disadvantaged groups. Additionally, it employs hyperedges to connect disadvantaged groups, promoting mutual learning and enhancing their node representations. Finally, dynamic weight adjustment is introduced to focus more on disadvantaged groups. Extensive experiments on real-world datasets demonstrate that FairHGNN outperforms state-of-the-art baselines, achieving superior fairness and utility (FairHGNN: <a href="https://github.com/Yang7777Lau/FairHGNN">https://github.com/Yang7777Lau/FairHGNN</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

FairHGNN: toward label-aware fairness in Heterogeneous Graph Neural Networks

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Yangqi Liu,
  • Xuemin Wang,
  • Chuangying Zhu,
  • Liang Chang,
  • Tianlong Gu

摘要

Heterogeneous Graph Neural Networks (HGNNs) excel in learning representations of Heterogeneous Graphs (HG) with rich structural and semantic information. However, many existing HGNN models tend to focus more on advantaged groups with abundant labeled data, while overlooking disadvantaged groups with fewer labels. This imbalance leads to the feature propagation process being dominated by the advantaged groups, which causes the disadvantaged groups to lose their unique characteristics during neighborhood aggregation. To address this fairness issue, termed label-aware fairness, we propose a novel framework, FairHGNN (Label-aware Fair Heterogeneous Graph Neural Network). FairHGNN leverages knowledge distillation to transfer neighborhood information from advantaged to disadvantaged groups. Additionally, it employs hyperedges to connect disadvantaged groups, promoting mutual learning and enhancing their node representations. Finally, dynamic weight adjustment is introduced to focus more on disadvantaged groups. Extensive experiments on real-world datasets demonstrate that FairHGNN outperforms state-of-the-art baselines, achieving superior fairness and utility (FairHGNN: https://github.com/Yang7777Lau/FairHGNN).