Graph Anomaly Detection Boundary Learning via Local and Global Structure Modeling

Graph anomaly detection aims to identify graphs within the graph set that differ from other graphs in structures and attributes, also called graph-level anomaly detection. Existing methods rely on graph neural networks to capture local structures and often fail to separate normal and abnormal graphs by a discriminative detection boundary. In this paper, we propose a Graph Anomaly Detection Boundary Learning framework (GADBL) by combining a standard graph view and a hypergraph view to capture both local and global structures. In addition, a new dual normalizing flow training module is designed to learn a comprehensive normal graph representation distribution. Finally, to better recognize anomalous graphs, we propose an anomaly detection boundary learning module under the optimization loss by introducing anomalous graph information. Extensive experiments on eight benchmarks show that GADBL outperforms eight baselines and its effectiveness is verified by ablation study and parameter analysis.