Graph anomaly detection based on hybrid node representation learning.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Anomaly detection in graph data is crucial for academia and industry.
• Graph Neural Networks (GNNs) show promise but struggle with class and semantic inconsistencies.
• Existing methods fail to fully resolve these inconsistencies, limiting performance.

Purpose of the Study:

• To develop a novel framework for graph anomaly detection that overcomes class and semantic inconsistencies.
• To improve the performance and reliability of anomaly detection models on graph data.

Main Methods:

• A novel framework combining a semantic fusion-based node representation module and an attention mechanism-based node representation module.
• The semantic fusion module utilizes Chebyshev polynomial graph filtering to capture signal components.
• The attention module adaptively learns node importance to address semantic inconsistency.

Main Results:

• The proposed method effectively captures high-frequency and low-frequency graph signal components.
• Adaptive learning of node importance significantly improves model performance.
• Experiments on five real-world datasets demonstrate superior performance over state-of-the-art methods.

Conclusions:

• The developed framework successfully resolves class and semantic inconsistencies in graph anomaly detection.
• The novel modules enhance the ability to detect anomalies in complex graph structures.
• The proposed method represents a significant advancement in GNN-based anomaly detection.