Rich Network Anomaly Detection Using Multivariate Data

Telecommunication networks are designed for high reliability but, given their complexity, when problems do occur they are difficult to detect and diagnose. Anomaly detection approaches typically provide cryptic results, resulting in extensive human effort for diagnosis. Using data from a 4G network, we focus on non-parametric change detection algorithms for anomaly detection and evaluate the performance of the algorithms with two variables: procedure duration and percent failing events. When an anomaly is detected, visual analytics are applied to infer the cause. The impact of our work is the proactive detection and cause analysis of anomalies and significantly reducing the number of dropped and degraded calls and sessions (9% to 27%).