ARISE: Graph Anomaly Detection on Attributed Networks via Substructure Awareness.

Area of Science:

• Data Mining
• Machine Learning
• Network Analysis

Background:

• Graph anomaly detection on attributed networks is a growing research area.
• Existing methods often overlook collective anomalous behavior within dense substructures.
• Distinguishing between attribute and topology anomalies remains a challenge.

Purpose of the Study:

• To propose a new framework, ARISE (Anomaly Recognition via substructure awareness), for graph anomaly detection on attributed networks.
• To improve topology anomaly detection by recognizing collective patterns in dense substructures.
• To simultaneously detect both topology and attribute anomalies.

Main Methods:

• ARISE employs a region proposal module to identify high-density substructures as suspicious regions.
• It calculates topology anomaly degree based on average node-pair similarity within substructures.
• A graph contrastive learning scheme is integrated to refine node attribute embeddings and detect attribute anomalies.

Main Results:

• ARISE demonstrates significant improvements in detection performance compared to state-of-the-art attributed network anomaly detection (ANAD) algorithms.
• The framework achieved up to 7.30% gains in Area Under the Curve (AUC).
• Substantial improvements were also observed in Area Under the Precision-Recall Curve (AUPRC), reaching up to 17.46%.

Conclusions:

• ARISE effectively detects both topology and attribute anomalies in attributed networks.
• Focusing on substructure awareness significantly enhances graph anomaly detection capabilities.
• The proposed framework offers a promising advancement in attributed network anomaly detection.