Dual-view graph-of-graph representation learning with graph Transformer for graph-level anomaly detection

Area of Science:

• Graph representation learning
• Machine learning
• Data mining

Background:

• Graph-Level Anomaly Detection (GLAD) aims to identify unusual graphs in datasets.
• Current Graph Neural Network (GNN) methods have limitations in capturing comprehensive intra-graph information and inter-graph relationships.

Purpose of the Study:

• To propose a novel unsupervised GLAD method addressing limitations of existing GNN-based approaches.
• To enhance the detection of anomalies by considering both intra-graph and inter-graph perspectives.

Main Methods:

• Introduced a Graph Transformer to expand the receptive field for richer intra-graph feature extraction.
• Developed a Graph-of-Graph-based dual-view representation learning network to model cross-graph dependencies.
• Utilized multi-perspective anomaly scores for comprehensive anomaly quantification.

Main Results:

• The proposed method effectively captures both attribute and structural information within and between graphs.
• Demonstrated superior performance in detecting anomalies across multiple benchmark datasets.
• The dual-view approach provides a more robust assessment of graph anomalies.

Conclusions:

• The novel Dual-View Graph-of-Graph Representation Learning Network significantly improves unsupervised GLAD.
• Considering both intra-graph and inter-graph perspectives is crucial for effective anomaly detection.
• The method offers a comprehensive approach to identifying anomalous graphs.