Cross-View Negative-Free Contrastive Learning for Graph Anomaly Detection with High-Order Structure Augmentation

Graph anomaly detection has practical applications in various fields, such as cyber security, financial evaluation and medical care. Recently, contrastive-based and generative-based detection frameworks have achieved remarkable performance improvements. However, most of the existing paradigms overlook the drawback that the GCN-based framework may unconsciously aggregate abnormal nodes with their neighborhood normal partners. Moreover, these detection algorithms lack attention to high-order structural information. These lead to a reduction in the distinction between normal nodes and their opponents. To bridge the gaps above, this paper proposes a cross-view negative-free contrastive learning utilizing high-order structure for graph anomaly detection (CNCL-GAD) in this paper. Especially, different from the existing single-view contrastive paradigm, this paper develops the high-order structure as the augmented view to introduce more global abnormality discrimination with multi-view contrastive learning for graph anomaly detection (GAD). Then, to mitigate the false-negative phenomenon of imbalanced data in GAD tasks where the majority of selected contrastive negative samples are normal subgraphs, this paper proposes the cross-view negative-free contrastive strategy to only pull the positive subgraphs’ pairs between two views as close as possible. Furthermore, this paper integrates intra-view node-subgraph contrastive modules, attribute reconstruction modules, and cross-view subgraph-subgraph contrastive modules to simultaneously obtain more distinctions on structure and attribute. The extensive experiments conducted on benchmark datasets show that the proposed method achieves competitive or even superior performance compared with existing competitors.