Correlation Information Enhanced Graph Anomaly Detection via Hypergraph Transformation.

Area of Science:

• Artificial Intelligence
• Data Science
• Network Analysis

Background:

• Graph anomaly detection (GAD) is crucial for applications like fraud detection and cybersecurity.
• Graph neural networks (GNNs) are powerful for GAD but struggle with anomaly camouflage, where anomalies resemble normal data.
• Existing GNN models face challenges in distinguishing subtle anomalies due to feature similarity.

Purpose of the Study:

• To propose a novel approach, Correlation Information Enhanced Graph Anomaly Detection (CIE-GAD), to overcome anomaly camouflage in GAD.
• To enhance the ability of GNNs to detect anomalies by effectively utilizing correlation information among graph edges.
• To improve the performance and robustness of GAD methods on complex, real-world graph datasets.

Main Methods:

• Constructing a hypergraph to model co-occurrence relationships between adjacent edges, capturing subtle differences between normal and abnormal samples.
• Developing a spectral convolution mechanism with node-level attention fusion to capture multifrequency signals and mitigate local heterophily.
• Enhancing the extraction of sample correlation information to counteract feature similarity caused by anomaly camouflage.

Main Results:

• CIE-GAD significantly outperforms state-of-the-art GAD methods on diverse real-world datasets.
• Achieved up to 3.47% improvement in AUC-PR (Area Under the Precision-Recall Curve), with an average gain of 1.5%.
• Demonstrated effectiveness in detecting anomalies even when they exhibit feature similarity with normal instances.

Conclusions:

• The proposed CIE-GAD effectively addresses the challenge of anomaly camouflage in graph anomaly detection.
• The novel hypergraph construction and spectral convolution mechanism enhance GNNs' capability for robust GAD.
• CIE-GAD offers a promising advancement for detecting anomalies in complex graph-structured data across various applications.