Deep graph level anomaly detection with contrastive learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Graph level anomaly detection (GLAD) is crucial for identifying unusual graph structures and features.
• Existing methods struggle with comprehensive graph representations, effective anomaly evaluation, and imbalanced data.

Purpose of the Study:

• To develop an end-to-end framework for GLAD addressing key challenges.
• To enhance graph-level representations for distinguishing normal and anomalous graphs.
• To introduce a novel anomaly evaluation paradigm for local and global graph perspectives.

Main Methods:

• Combines graph neural networks (GNNs) and contrastive learning.
• Employs a graph convolution autoencoder with a perturbed graph encoder.
• Enhances graph representations at both node and graph levels using contrastive strategies.

Main Results:

• The proposed GLAD framework effectively detects anomalies in graph datasets.
• Achieved superior performance compared to existing advanced methods across ten real-world datasets.
• Demonstrated effectiveness in molecular, protein, and social network anomaly detection.

Conclusions:

• The integrated GNN and contrastive learning approach provides a robust solution for GLAD.
• The novel evaluation method successfully captures graph anomalies from multiple perspectives.
• The framework offers significant advancements in distinguishing anomalous graphs in various domains.