Graph Anomaly Detection via Multiscale Contrastive Self-Supervised Learning From Local to Global

Graph anomaly detection is a challenging task in graph data mining, aiming to recognize unconventional patterns within a network. Recently, there has been increasing attention on graph anomaly detection based on contrastive learning due to its high adaptability to the sample imbalance problem. However, most existing work typically focuses on the contrast of local views while neglecting global comparison information, leading to suboptimal performance. To address this issue, we introduce a new multiscale contrastive self-supervised learning framework for graph anomaly detection (GADMCLG). Our approach incorporates local-level contrasts involving node-node and node- subgraph contrast, and global-level subgraph-subgraph contrast. The former mines localized abnormal information, while the latter is intended to capture global anomalous patterns. Specifically, our proposed subgraph-subgraph contrast adopts the h-order neighbor subgraph sampling instead of augmented subgraphs through edge perturbation. This sampling strategy ensures a comprehensive observation of the neighborhood surrounding the target node, thereby mitigating the introduction of extraneous noise and providing interpretability for the detected results. Furthermore, we incorporate a subgraph centralization technique to reduce the bias caused by the absolute position of subgraphs in the attribute space, which enhances the model's ability to identify anomalies at different scales. Extensive experimental results on six real-world datasets demonstrate the effectiveness of our method and its superiority compared with state-of-the-art approaches.