Frequency Self-Adaptation Graph Neural Network for Unsupervised Graph Anomaly Detection

Abstract

Unsupervised Graph Anomaly Detection (UGAD) seeks to identify abnormal patterns in graphs without relying on labeled data. Among existing UGAD methods, Graph Neural Networks (GNNs) have played a critical role in learning effective representation for detection by filtering low-frequency graph signals. However, the presence of anomalies can shift the frequency band of graph signals toward higher frequencies, thereby violating the fundamental assumptions underlying GNNs and anomaly detection frameworks. To address this challenge, the design of novel graph filters has garnered significant attention, with recent approaches leveraging anomaly labels in a semi-supervised manner. Nonetheless, the absence of anomaly labels in real-world scenarios has rendered these methods impractical, leaving the question of how to design effective filters in an unsupervised manner largely unexplored. To bridge this gap, we propose a novel Frequency Self-Adaptation Graph Neural Network for Unsupervised Graph Anomaly Detection (FAGAD). Specifically, FAGAD adaptively fuses signals across multiple frequency bands using full-pass signals as a reference. It is optimized via a self-supervised learning approach, enabling the generation of effective representations for unsupervised graph anomaly detection. Experimental results demonstrate that FAGAD achieves state-of-the-art performance on both artificially injected datasets and real-world datasets. The code and datasets are publicly available at https://github.com/eaglelab-zju/FAGAD.