B1142
Title: Online change point detection with adaptive learning for multivariate processes
Authors: Konstantinos Bourazas - University of Cyprus (Cyprus) [presenting]
Konstantinos Fokianos - University of Cyprus (Cyprus)
Christos Panayiotou - University of Cyprus (Cyprus)
Marios Polycarpou - University of Cyprus (Cyprus)
Abstract: Online anomaly detection for multivariate data is of great interest in a variety of fields. In order to achieve efficient detection, developed methodologies typically rely on assumptions that may not be met in practice. In this work, our objective is to provide an effective detection scheme for multivariate processes while relaxing the initial requirements as much as possible. We develop a CUSUM-type procedure for detecting persistent changes to the mean vector of the data in an online fashion. The proposed methodology is distribution-free, and it is based on the Kernel Density Estimation. In addition it is self-starting, and adaptive in relation to the magnitude, or direction of a change. Furthermore, a post-alarm estimate for the change point location and the size of the shift detected is available. A simulation study evaluates its performance against standard alternative for various anomaly scenarios, while two applications to real data demonstrate its use in practice.
