GADACE: Graph Anomaly Detection Combining Attribute Contrast and Structure Reconstruction

Unsupervised graph anomaly detection aims to identify nodes that deviate from typical behaviors in graphs. Existing approaches can be briefly categorized into two main groups, namely, reconstruction-based approaches that detect anomalies through reconstruction errors, and contrastive learning-based approaches that focus on local differences to identify inconsistency. Nevertheless, neither of these approaches fully leverages the graph’s information, leading to suboptimal performance. In this paper, we present GADACE, a novel framework that integrates local attribute contrast with global structure reconstruction to generate comprehensive scores for graph anomaly detection. In this framework, a multi-level, cross-view contrastive network based on MLPs is utilized to capture local inconsistency. Meanwhile, MLP-based autoencoders are trained on both original and diffusion-augmented features to improve link prediction and global inconsistency identification. Then, the framework assigns an anomaly score to each node based on the local and global inconsistencies. Extensive experimental results on real-world datasets verify the effectiveness of our approach.