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

Abstract

Unsupervised Graph Anomaly Detection (UGAD) aims to identify abnormal patterns within a graph without supervision. 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, anomalies can shift the frequency band of the graph signal towards the high-frequency area, disrupting the fundamental assumptions of GNNs and anomaly detection based on them. To tackle this challenge, designing new graph filters has attracted increasing attention and a few efforts have been made recently under the supervision of anomaly labels. Nevertheless, the lack of anomaly labels has caused the semi-supervised methods to fail in practice, and the question of how to design proper filters in an unsupervised manner remains unexplored. To bridge this gap, in this paper, we propose a novel Frequency Self-Adaptation Graph Neural Network for Unsupervised Graph Anomaly Detection (FAGAD). Specifically, the \model is able to adaptively fuse signals from multiple frequency bands by taking full-pass signals as a reference. It is optimized in a self-supervised manner and produces effective representations for unsupervised graph anomaly detection. Experimental results demonstrate that the FAGAD achieves state-of-the-art performance on both artificially injected datasets and real-world datasets. Our code and datasets are open source at https://anonymous.4open.science/r/FAGAD.

Keywords: Unsupervised Graph Anomaly Detection, Graph Neural Networks, Graph Filtering