Multi-Frequency Aware Reconstruction Model For Unsupervised Graph Anomaly Detection

Xinyi Wan, Pengda Li, Yihong Huang, Liping Wang

Abstract: Unsupervised graph anomaly detection (UGAD) is defined as identifying abnormal nodes that deviate from the normal distribution in a graph without any labels, and is widely used in financial fraud detection and spam users detection in social networks. However, in real-world datasets, the low proportion of abnormal nodes leads to the prevalence of heterophily edges in the graph. This makes abnormal and normal nodes increasingly indistinguishable during the neighbor information aggregation, thus leading to feature-smoothing in traditional UGAD models, resulting in suboptimal anomaly detection performance. Therefore, the key challenge in UGAD is to prevent node feature-smoothing from heterophily edges without labeled data, for better differentiation between normal and anomalous nodes.To address this challenge, we observed that an increase in abnormal nodes correlates with stronger high-frequency signals. And we propose MFA-UGAD (Multi-Frequency Aware reconstruction model for Unsupervised Graph Anomaly Detection), a Beta Wavelet kernel-based model. MFA-UGAD employs encoder with multi-spectral localized band-pass filters to capture multi-order neighborhoods and multi-frequency signals, effectively detecting both global and local anomalies. Additionally, we introduce a novel spectral reconstruction loss to amplify abnormal high-frequency signals and mitigate node information feature-smoothing. Experiments on five real-world datasets validate the effectiveness of our approach, achieving on average an increase of over 7.932% in AUROC compared with the top competitors.

External IDs:dblp:conf/ijcnn/WanLHW25