Scholarlink

<p>Dynamic graph anomaly detection aims to identify irregular patterns in evolving graph data. Conventional GNN-based methods often struggle with over-smoothing, over-squashing, and fail to effectively capture temporal dependencies. To address these limitations, we propose GLT-GAD, a Global-Local Temporal Graph Transformer designed to enhance anomaly detection in dynamic graphs. GLT-GAD introduces a hierarchical Transformer framework that reduces computational complexity while preserving global dependencies by partitioning the graph into manageable clusters and modeling intra-cluster and inter-cluster dependencies. To incorporate structural inductive biases, we construct ego-graph for each node to capture local structural information. Furthermore, GLT-GAD integrates time tokens within node sequences to model temporal dependencies and improve anomaly detection performance in evolving graph structures. Extensive experiments on four public benchmark datasets demonstrate that GLT-GAD achieves competitive performance compared with existing methods, indicating its effectiveness in dynamic graph anomaly detection. Code is available at <a href="https://github.com/senllh/GLT-GAD.">https://github.com/senllh/GLT-GAD.</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

GLT-GAD: Global-Local Temporal Graph Transformer for Anomaly Detection in Dynamic Graphs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月28日

查看全文
  • Hanbin Lu,
  • Haosen Wang

摘要

Dynamic graph anomaly detection aims to identify irregular patterns in evolving graph data. Conventional GNN-based methods often struggle with over-smoothing, over-squashing, and fail to effectively capture temporal dependencies. To address these limitations, we propose GLT-GAD, a Global-Local Temporal Graph Transformer designed to enhance anomaly detection in dynamic graphs. GLT-GAD introduces a hierarchical Transformer framework that reduces computational complexity while preserving global dependencies by partitioning the graph into manageable clusters and modeling intra-cluster and inter-cluster dependencies. To incorporate structural inductive biases, we construct ego-graph for each node to capture local structural information. Furthermore, GLT-GAD integrates time tokens within node sequences to model temporal dependencies and improve anomaly detection performance in evolving graph structures. Extensive experiments on four public benchmark datasets demonstrate that GLT-GAD achieves competitive performance compared with existing methods, indicating its effectiveness in dynamic graph anomaly detection. Code is available at https://github.com/senllh/GLT-GAD.