Scholarlink

<p>Graph neural networks (GNNs) have emerged as powerful tools for addressing graph-structured data, demonstrating superior performance across diverse applications. However, their “black box” nature obscures the reasoning behind predictions, limiting trust and practical deployment. We propose NXGNN (<b>N</b>earest-neighbor e<b>X</b>plainable <b>G</b>raph <b>N</b>eural <b>N</b>etwork), an interpretability framework that combines nearest neighbor estimation with neural feature transformation to generate explanatory subgraphs, aiming to align the model’s subgraph predictions closely with those of the original graph. Through extensive experiments on both synthetic datasets (BA-Shapes, Tree-Grid) and real-world networks (Cora, CiteSeer), NXGNN demonstrates superior performance compared to existing explainers while maintaining predictive accuracy. By providing a method that balances high predictive power with high interpretability, NXGNN contributes one more method to the development of trustworthy artificial intelligence systems.</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

Explainability graph neural networks with nearest neighbor estimate interpretations

  • 期刊论文

  • 发表日期：

    2025年5月31日

查看全文
  • Duy Nguyen,
  • Thanh Le,
  • Bac Le

摘要

Graph neural networks (GNNs) have emerged as powerful tools for addressing graph-structured data, demonstrating superior performance across diverse applications. However, their “black box” nature obscures the reasoning behind predictions, limiting trust and practical deployment. We propose NXGNN (Nearest-neighbor eXplainable Graph Neural Network), an interpretability framework that combines nearest neighbor estimation with neural feature transformation to generate explanatory subgraphs, aiming to align the model’s subgraph predictions closely with those of the original graph. Through extensive experiments on both synthetic datasets (BA-Shapes, Tree-Grid) and real-world networks (Cora, CiteSeer), NXGNN demonstrates superior performance compared to existing explainers while maintaining predictive accuracy. By providing a method that balances high predictive power with high interpretability, NXGNN contributes one more method to the development of trustworthy artificial intelligence systems.