Scholarlink

The rapid advancement of Graph Neural Networks (GNNs) has brought model interpretability to the forefront of machine learning research, especially in high-stakes domains such as healthcare, finance, and legal applications. Despite their impressive performance across various tasks, GNNs remain inherently opaque, with their decision-making processes often obscured by their “black-box” nature. This lack of transparency hinders trust and limits their real-world adoption in critical scenarios. To bridge this gap, we present AMSExplainer, a novel interpretable GNN framework built upon an Adaptive Multi-Scale Variational Graph Auto-Encoder (AMSVGAE) architecture. The proposed AMSVGAE component employs a sophisticated feature aggregation mechanism that operates across multiple neighborhood scales, enabling comprehensive capture of both local structural patterns and global graph dependencies. Furthermore, AMSExplainer incorporates an innovative causal disentanglement module that systematically separates fundamental causal factors from non-essential features, thereby offering unprecedented insights into the causal relationships driving model predictions. Designed with model-agnostic principles, our framework maintains compatibility with arbitrary GNN architectures without requiring access to their internal parameters or computational graphs. Extensive experimental evaluations demonstrate that AMSExplainer consistently generates explanations that are not only more accurate but also more compact and faithful to the model’s actual reasoning process compared to existing approaches.

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

AMSVGAE-Based Causal Inference for Interpretable Graph Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jixuan Wu,
  • Limei Lin,
  • Yanze Huang,
  • Xiaoding Wang,
  • Jianxi Fan,
  • Xiaohua Jia

摘要

The rapid advancement of Graph Neural Networks (GNNs) has brought model interpretability to the forefront of machine learning research, especially in high-stakes domains such as healthcare, finance, and legal applications. Despite their impressive performance across various tasks, GNNs remain inherently opaque, with their decision-making processes often obscured by their “black-box” nature. This lack of transparency hinders trust and limits their real-world adoption in critical scenarios. To bridge this gap, we present AMSExplainer, a novel interpretable GNN framework built upon an Adaptive Multi-Scale Variational Graph Auto-Encoder (AMSVGAE) architecture. The proposed AMSVGAE component employs a sophisticated feature aggregation mechanism that operates across multiple neighborhood scales, enabling comprehensive capture of both local structural patterns and global graph dependencies. Furthermore, AMSExplainer incorporates an innovative causal disentanglement module that systematically separates fundamental causal factors from non-essential features, thereby offering unprecedented insights into the causal relationships driving model predictions. Designed with model-agnostic principles, our framework maintains compatibility with arbitrary GNN architectures without requiring access to their internal parameters or computational graphs. Extensive experimental evaluations demonstrate that AMSExplainer consistently generates explanations that are not only more accurate but also more compact and faithful to the model’s actual reasoning process compared to existing approaches.