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

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Unsupervised Graph Anomaly Detection (UGAD) methods often rely on Graph Neural Networks (GNNs) that filter low-frequency graph signals.
• Anomalies can shift graph signal frequencies to higher bands, violating GNN assumptions and hindering detection.
• Existing advanced graph filters often require anomaly labels, limiting their real-world applicability.

Purpose of the Study:

• To address the limitations of current unsupervised graph anomaly detection methods.
• To propose a novel approach for designing effective graph filters in an unsupervised manner.
• To develop a Graph Neural Network capable of handling frequency shifts caused by anomalies.

Main Methods:

• Proposing the Frequency Self-Adaptation Graph Neural Network (FAGAD).
• Adaptively fusing graph signals across multiple frequency bands using full-pass signals as a reference.
• Optimizing the model via a self-supervised learning approach for representation generation.

Main Results:

• FAGAD demonstrates state-of-the-art performance on anomaly detection tasks.
• The method achieves high accuracy on both artificially generated and real-world datasets.
• The proposed approach effectively handles the challenges posed by frequency shifts in graph signals.

Conclusions:

• FAGAD offers a robust solution for unsupervised graph anomaly detection.
• The self-supervised learning framework enables effective representation learning without labeled data.
• The adaptive fusion of multi-frequency signals is key to FAGAD's superior performance.