Nsgad: A Neighborhood Subtraction-Based Framework for Unsupervised Graph Anomaly Detection

Abstract

Graph anomaly detection is an essential task in artificial intelligence. With the emergence of Graph Neural Networks (GNNs), significant progress has been made in graph anomaly detection. However, most GNNs use Graph Convolutional Networks (GCN) as their backbone, which often filters out high-frequency signals crucial for anomaly detection, making anomalies difficult to detect. To address this issue, we propose a Neighborhood Subtraction-based framework for unsupervised Graph Anomaly Detection (NSGAD). This framework ingeniously utilizes neighborhood subtraction to amplify high-frequency signals, thereby magnifying abnormal information in the latent space. Additionally, we introduce a low-frequency reconstruction module to preserve the integrity of low-frequency information. The experimental results on synthetic and real anomaly datasets demonstrated that NSGAD outperforms the latest state-of-the-art methods by an average of 3% in terms of Area Under Curve (AUC) while simultaneously reducing the temporal cost by around 65%. Our code and dataset will be available athttps://github.com/xxx/xxx.

Keywords: Graph anomaly detection, Unsupervised learning, Neighborhood subtraction, Low-frequency reconstruction