A New Approach for Outlier Detection in Near Real Time
Julius Eiweck, Colin Pattinson, Reinhold Behringer, Alexander K. Seewald · 2010
Outlier detection methods have been suggested for a broad range of applications. They are obligatory in different monitoring tasks (e.g. mobile phone monitoring, credit card usage monitoring). Aim is the detection of sudden changes in the usage pattern which may indicate deviations from normal usage. Parametric (statistical) and non parametric methods combined with univariate and multivariate methods form most of the body of research in anomaly detection. In this paper a novel approach for outlier detection is proposed. For the validation of the new approach we have applied the method on voice traffic time-series obtained from a real-life mobile network. Main focus is the incorporation of standard statistical methods, the ability to implement specific heuristics derived from domain expert knowledge and near real-time performance. The reliability achieved with this novel approach for outlier detection is sufficient to be used in operational processes. Enabling the detection of outliers in near real time supports a broad range of monitoring tasks to be performed by the operational staff.