Unified representation and scoring framework for anomaly detection in attributed networks with emphasis on structural...

Area of Science:

• Computer Science
• Data Science
• Network Analysis

Background:

• Anomaly detection in attributed networks is crucial for security and fraud detection.
• Existing methods struggle with anomalies stemming from both structure and attributes.
• Single learning paradigms limit the capture of diverse anomalous behaviors.

Purpose of the Study:

• To propose a comprehensive hybrid framework for attributed network anomaly detection.
• To integrate multiple learning components for robust anomaly identification.
• To improve the detection of anomalies arising from structural and attribute inconsistencies.

Main Methods:

• A hybrid framework integrating graph structure reconstruction and attribute reconstruction.
• Community-aware contrastive learning for discriminative representation.
• Similarity-aware anomaly scoring for neighborhood-based refinement.
• Evaluation on six benchmark datasets: BlogCatalog, Flickr, ACM, Cora, Citeseer, and Pubmed.

Main Results:

• The proposed framework significantly outperforms state-of-the-art baselines.
• Achieved superior performance across AUC and AUPR evaluation metrics.
• Ablation studies confirmed the contribution of each module.
• Parameter sensitivity analysis demonstrated framework robustness.

Conclusions:

• The hybrid framework effectively detects a wide range of anomalies in complex attributed graphs.
• The unified design captures global structural patterns and local semantic consistencies.
• The method demonstrates effectiveness and generalizability in attributed network anomaly detection.