Anomaly Detection on Attributed Network Based on Hyperbolic Radial Distance
Tianyu Liu, Ruixin Yan, Irwin King · 2024
Anomaly detection in attributed networks is crucial for numerous real-world applications, including cybersecurity, social network analysis, and bioinformatics. Traditional anomaly detection methods face two key limitations: reliance on Euclidean space for node embeddings and the assumption of a clear geometric separation between normal instances and anomalies. This paper introduces an innovative anomaly detection approach using hyperbolic graph neural networks (GNNs) to overcome these challenges. Firstly, our unsupervised model utilizes node embeddings within hyperbolic space, adept at representing hierarchical and complex network structures. Secondly, we introduce a novel anomaly detection metric based on hyperbolic radial distance, effectively identifying anomalies without requiring distinct separation in the feature space. Extensive experiments demonstrate our model’s enhanced performance over traditional methods, highlighting its potential in addressing the intricacies and limitations of anomaly detection in complex network environments.