SAMCL: Subgraph-Aligned Multiview Contrastive Learning for Graph Anomaly Detection.

Area of Science:

• Graph theory
• Machine learning
• Data mining

Background:

• Graph anomaly detection (GAD) is crucial for networks like social and financial systems.
• Graph contrastive learning (GCL) is a leading GAD approach, but often overlooks subgraph-subgraph comparisons.
• Existing GCL methods struggle with non-aligned subgraph pairs due to size or node index differences.

Purpose of the Study:

• To propose a novel subgraph-aligned multiview contrastive approach (SAMCL) for graph anomaly detection.
• To address the limitation of existing GCL methods by introducing subgraph-subgraph level contrast.
• To overcome the 'nonaligned' issue in subgraph pair comparison for accurate similarity measurement.

Main Methods:

• Generated multiview augmented subgraphs by capturing diverse neighbors of target nodes.
• Developed a subgraph-aligned strategy using Earth Mover's Distance (EMD) for non-aligned subgraph similarity, considering node embeddings and topology.
• Integrated subgraph-aligned contrastive learning, intra-view node-subgraph contrastive learning, and masked subgraph reconstruction for anomaly scoring.

Main Results:

• SAMCL effectively addresses the subgraph-subgraph contrastive-level gap in GAD.
• The proposed subgraph-aligned strategy accurately measures similarity between non-aligned subgraph pairs.
• Experiments show significant performance gains, up to 6.36% improvement on the ACM dataset, compared to state-of-the-art methods.

Conclusions:

• SAMCL offers a novel and effective approach to graph anomaly detection by incorporating subgraph-subgraph contrast.
• The method successfully handles non-aligned subgraph comparisons, enhancing anomaly detection accuracy.
• The combined contrastive learning and reconstruction modules provide a robust framework for identifying anomalies in graph data.