Scholarlink

<p>As an indispensable pillar of responsible AI systems, interpretability studies on Graph Neural Networks (GNNs) have achieved breakthroughs that directly contribute to the robustness of intelligent decision-making. However, most existing interpretability methods focus primarily on homogeneous graphs and fall short when it comes to interpreting heterogeneous data in real-world scenarios, such as financial network and social network. Addressing interpretation challenges in more complex heterogeneous graphs has emerged as a critical issue. To better leverage the rich information contained in heterogeneous graphs and enhance the safety and reliability of model deployment, we introduce HGExplainer, a novel interpreter based on meta-path perturbation for heterogeneous graph neural networks. HGExplainer transforms heterogeneous data into a more manageable homogeneous format through pre-defined meta-paths. It then employs a perturbation strategy to identify the key components that influence downstream tasks. This approach not only enhances interpretability but also ensures that models can be used more safely and reliably. Our experimental results on public datasets demonstrate the effectiveness and interpretability of HGExplainer, highlighting its potential to improve the trustworthiness and practical applicability of GNN models in real-world applications.</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

HGExplainer: Toward Interpretable Heterogeneous Graph Neural Networks via Meta-path Perturbation

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Yanjing Wang,
  • Zhuohan Zhang,
  • Jialiang Yin,
  • Xintong Li,
  • Bin Shi,
  • Bo Dong

摘要

As an indispensable pillar of responsible AI systems, interpretability studies on Graph Neural Networks (GNNs) have achieved breakthroughs that directly contribute to the robustness of intelligent decision-making. However, most existing interpretability methods focus primarily on homogeneous graphs and fall short when it comes to interpreting heterogeneous data in real-world scenarios, such as financial network and social network. Addressing interpretation challenges in more complex heterogeneous graphs has emerged as a critical issue. To better leverage the rich information contained in heterogeneous graphs and enhance the safety and reliability of model deployment, we introduce HGExplainer, a novel interpreter based on meta-path perturbation for heterogeneous graph neural networks. HGExplainer transforms heterogeneous data into a more manageable homogeneous format through pre-defined meta-paths. It then employs a perturbation strategy to identify the key components that influence downstream tasks. This approach not only enhances interpretability but also ensures that models can be used more safely and reliably. Our experimental results on public datasets demonstrate the effectiveness and interpretability of HGExplainer, highlighting its potential to improve the trustworthiness and practical applicability of GNN models in real-world applications.