Enhancing Multi-view Contrastive Learning for Graph Anomaly Detection

Qingcheng Lu, Nannan Wu, Yiming Zhao, Wenjun Wang, Quannan Zu

Abstract: Graph anomaly detection (GAD) has garnered considerable attention owing to its wide-ranging applications in real-world scenarios, including financial network and cybersecurity. Due to the difficulty associated with learning intricate connectivity and high-dimensional attributes in graphs, GAD task poses significant challenges. Existing deep learning methods extract local information between the target node and its neighbors for GAD. Nevertheless, they neglect the global information in the graph which results in suboptimal performance. In this paper, we propose a self-supervised learning framework integrating contrastive learning module and attribute reconstruction module, which can effectively capture both local and global information for GAD. Specifically, we regard original graph as local view, and global view is generated based on this local view. In contrastive learning module, target node and subgraph are sampled from both views to construct multi-scale contrast combinations. With intra- and intre- view contrastive strategy, we can effectively capture both local and global information of anomaly nodes. In attribute reconstruction module, masked graph auto-encoder is employed to further capture anomaly information within the attribute space. Extensive experiments are conducted on six benchmark datasets and results reveal that our approach outperforms the current state-of-the-art methods.