GAD-NR: Graph Anomaly Detection via Neighborhood Reconstruction.

Area of Science:

• Computer Science
• Data Science
• Network Security

Background:

• Graph Anomaly Detection (GAD) identifies abnormal nodes in graphs for applications like fraud and spam detection.
• Graph Auto-Encoders (GAEs) are common GAD methods but struggle with complex structural anomalies due to focus on direct link reconstruction.

Purpose of the Study:

• To introduce GAD-NR, a novel Graph Auto-Encoder variant for enhanced Graph Anomaly Detection.
• To improve the detection of complex structural anomalies that existing GAE models fail to identify.

Main Methods:

• GAD-NR incorporates neighborhood reconstruction, capturing local structure, self-attributes, and neighbor attributes.
• Anomaly detection is achieved by comparing neighborhood reconstruction loss between normal and anomalous nodes.

Main Results:

• GAD-NR demonstrated significant improvements, up to 30% increase in AUC, over state-of-the-art competitors on six real-world datasets.
• Unlike existing methods that detect only one or two anomaly types, GAD-NR effectively detects all three studied anomaly types.

Conclusions:

• GAD-NR offers a robust and comprehensive solution for Graph Anomaly Detection, outperforming current techniques.
• The open availability of GAD-NR's source code facilitates further research and application in diverse domains.