Graph Anomaly Detection via Multi-View Discriminative Awareness Learning

With the deeper research on attributed networks, graph anomaly detection is becoming an increasingly important topic. It aims to identify patterns deviating from a majority of nodes. Currently, graph anomaly detection algorithms based on reconstruction-based learning and contrastive-based learning have gained significant attention. To harness diverse supervised signals, an intuitive approach is to find an elegant strategy to fuse these two paradigms, forming the hybrid learning paradigm. Despite the success of the hybrid learning paradigm, due to its subgraph sampling based approach, it still grapples with issues related to unreliable neighborhood information and the neglect of topological details. To address these limitations, this paper proposes a new hybrid learning paradigm via multi-view discriminative awareness learning for graph anomaly detection. Unlike the previous hybrid learning paradigm, the graph reconstruction module fully incorporates attribute and topology information, enhancing the comprehensiveness of data reconstruction. Moreover, the multi-view discrimination module employs a view-level contrast method based on the complete graph, which helps to comprehensively extract the information in the attributed network and mitigates the neighborhood unreliability without increasing the complexity. The experimental results, obtained from a rigorous evaluation on six benchmark datasets, demonstrate the effectiveness of the proposed method compared to existing baseline methods.