CGAD: A novel contrastive learning-based framework for anomaly detection in attributed networks

Abstract

The efficient detection of anomalies in attributed networks is crucial for various applications. Although contrastive learning methods have shown potential in this domain, most existing studies still suffer from issues such as negative sampling bias and limited utilization of neighborhood information. Moreover, these methods often fail to account for the impact of systematic noise. In this paper, we introduce a purpose-built framework called Contrastive learning-based Graph Anomaly Detection (CGAD) to address these specific challenges. In CGAD, we present a debiased negative sampling approach, which intelligently selects negative nodes based on community distribution to mitigate negative sampling bias. Additionally, our approach incorporates multiple masking strategies to construct multi-scale instance pairs and leverages the Robust InfoNCE (RINCE) technique to effectively filter out noise, thus enhancing the overall performance of the framework. We evaluate the performance of CGAD across four diverse datasets and conduct a comparative analysis against seven baseline models. The experimental results consistently demonstrate the superior performance of CGAD.