Anomaly detection with dual-channel heterogeneous graph based on hypersphere learning

Graph anomaly detection is essential for identifying irregular patterns and outliers within complex network structures in domains like social networks, cybersecurity, finance, and transportation systems. It helps detect security breaches, fraud, and errors, improving decision-making and system reliability. Although current methodologies have advanced the unsupervised detection of graph anomalies, they frequently fail to fully address the nuanced specificities of graph anomalies, such as anomalous nodes, edges, and subgraphs. To overcome this limitation, the study presents MulDualGNN, a unique dual-channel heterogeneous graph anomaly detection framework. MulDualGNN incorporates a global substructure-aware GNN and a local substructure-aware GNN to capture both global and local substructure properties for accurate anomaly detection. Our model incorporates a multi-hypersphere learning target function, which includes macroscopic and mesoscopic hyperspheres. These hyperspheres measure abnormal nodes that deviate from most normal nodes in the entire graph and community structure, respectively. To overcome the model collapse problem in multi-hypersphere learning, our model utilizes the EmbSim similarity function to optimize the training target. The effectiveness and performance advantages of the proposed method are evaluated through extensive experiments on five datasets. The results demonstrate the superior performance of our approach in graph anomaly detection tasks.