Anomaly Detection in Dynamic Networks using Multi-view Time-Series Hypersphere Learning.

Detecting anomalous patterns from dynamic and multi-attributed network systems has been a challenging problem due to the complication of temporal dynamics and the variations reflected in multiple data sources. We propose a Multi-view Time-Series Hypersphere Learning (MTHL) approach that leverages multi-view learning and support vector description to tackle this problem. Given a dynamic network with time-varying edge and node properties, MTHL projects multi-view time-series data into a shared latent subspace, and then learns a compact hypersphere surrounding normal samples with soft constraints. The learned hypersphere allows for effectively distinguishing normal and abnormal cases. We further propose an efficient, two-stage alternating optimization algorithm as a solution to the MTHL. Extensive experiments are conducted on both synthetic and real datasets. Results demonstrate that our method outperforms the state-of-the-art baseline methods in detecting three types of events that involve (i) time-varying features alone, (ii) time-aggregated features alone, as well as (iii) both features. Moreover, our approach exhibits consistent and good performance in face of issues including noises, anomaly pollution in training phase and data imbalance.