Advancing unsupervised graph anomaly detection: A multi-level contrastive learning framework to mitigate local consistency deception
W. Wu, Yijun Gu · Neurocomputing · 2025
Graph anomaly detection plays a crucial role in various fields such as social networks, financial transactions, and cybersecurity. Due to the high cost of acquiring labeled node data, contrastive learning, an important paradigm in unsupervised learning, has gained significant attention and demonstrated effective performance in practical applications. Traditional contrastive learning methods evaluate anomalies by detecting inconsistencies between nodes and their neighbors or subgraphs, often expanding these scopes using techniques like Random Walk with Restart (RWR) to gather more information. However, in real-world anomaly datasets, normal and anomalous nodes often exhibit complex hybrid distributions. This complexity can deceive the consistency discrimination of contrastive learning methods under local views, leading to substantial performance drops. To address this issue, we propose TCL-GAD, a novel solution that incorporates global information for graph anomaly detection. TCL-GAD uses a Simple GCN module to extract node and neighborhood information while employing a Hierarchical Transformer to prevent over-reliance on higher-order node information, which could negatively impact the model, while efficiently extracting valuable global information. The Multi-Level Contrastive Learning framework enhances negative node sampling and effectively leverages multi-level graph information to boost overall detection performance. We evaluated TCL-GAD against thirteen state-of-the-art baseline methods across seven diverse datasets. Results show that TCL-GAD achieves excellent operational efficiency and consistently demonstrates superior performance, showing significant improvements over the runner-up methods especially on complex real-world datasets.