基于超球面对偶学习的双通道图异常检测方法

Graph anomaly detection, as a crucial data mining task, focuses on identifying anomalous nodes that significantly deviate from the majority of the nodes. With the advancement of unsupervised graph neural network techniques, various efficient methods have been developed to detect potential anomalies in graph data, including those based on density estimation and generative adversarial networks. However, these methods often focus on generating high-quality representations for unsupervised graph anomaly detection and tend to overlook the characteristics of graph anomalies. Consequently, this paper proposes a dual-channel heterogeneous graph anomaly detection model (HD-GAD). Its architecture includes two graph neural networks, i.e. a global substructure-aware GNN (Graph Neural Network) and a local substructure-aware GNN, designed to capture global and local substructural properties for graph anomaly detection. Additionally, the model introduces a multi-hypersphere learning (MHL) objective based on dual inference, which measures anomalies deviating from the overall graph/community structure from macro and meso hypersphere perspectives. The HD-GAD model utilizes the similarity function EmbSim to optimize the training objective, mitigating model collapse issues in multi-hypersphere learning. Comprehensive experiments conducted on five different datasets demonstrated that the AUC (Area Under Curve) values exceeded 0.9 in most cases, achieving industry-leading levels and further proving the HD-GAD model's efficiency and performance advantages in graph anomaly detection tasks.