Adapt Denoising-based Graph Anomaly Detection via Multi-Scale Contrastive Learning

Detecting anomalies in attributed networks has become increasingly important in practical applications such as cybersecurity and e-commerce. However, the presence of noise in the graph data often degrades the model’s performance by distorting the underlying patterns. To address this challenge, we propose a novel framework, Adaptive Denoising-based Graph Anomaly Detection (ADGAD), which integrates multi-scale contrastive learning and masked denoising techniques. This approach aims to reduce the impact of noise by usingf-supervised signals for noise filtering and applying an adaptive denoising strategy during training. The proposed method ensures more stable learning of normal patterns and improves the ability of the model to detect anomalous nodes. Experimental evaluations on six benchmark datasets demonstrate that ADGAD outperforms existing methods in terms of detection accuracy and stability, highlighting its effectiveness in noisy graph environments.