Guarding Graph Neural Networks for Unsupervised Graph Anomaly Detection.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Unsupervised graph anomaly detection identifies outliers in graph data without labels.
• Graph neural networks (GNNs) learn node representations by aggregating neighborhood information.
• Anomalies disrupt GNNs' assumption of neighborhood consistency, degrading performance.

Purpose of the Study:

• To address the adverse effects of graph anomalies on GNNs in unsupervised settings.
• To propose a novel framework, G3AD, for robust unsupervised graph anomaly detection.
• To improve node representation learning for enhanced anomaly identification.

Main Methods:

• G3AD employs auxiliary networks with correlation constraints to stabilize GNNs.
• An adaptive caching (AC) module prevents GNNs from reconstructing anomalous graph data.
• The framework is designed to be compatible with various GNN backbones.

Main Results:

• G3AD significantly outperforms 20 state-of-the-art methods on synthetic and real-world datasets.
• The proposed method demonstrates strong generalization ability across different GNN architectures.
• G3AD achieves superior performance in unsupervised graph anomaly detection tasks.

Conclusions:

• G3AD provides an effective solution for unsupervised graph anomaly detection by mitigating GNN vulnerabilities.
• The framework enhances the robustness and performance of GNNs in the presence of graph anomalies.
• G3AD represents a significant advancement in learning effective representations for anomaly detection in graphs.