Change Point Detection Based on Cluster Transition Distributions

Nowadays, Anomaly Detection (AD) models are incorporated into various systems, but they will become useless if they are not updated (i.e., re-trained) to keep up with changes in their external environment. When trying to automatically trigger AD model updates in response to environmental changes, one promising solution is considered to be the use of a Change Point Detection (CPD) method. Most existing methods impose stationary or Independent and Identically Distributed (IID) constraints on the target time-series, and therefore are not suitable for our target time-series in telecommunications that often fluctuate periodically. In this paper, we propose a new clustering-based CPD method for detecting changes in non-stationary time-series. The proposed method enables pattern changes to be detected by tracking cluster transitions and calculating the distance between the cluster transition distributions for the past and current periods. The accuracy of the proposed CPD method itself and the effect of re-training the AD model at the detected change points are shown by using real hourly time-series data at mobile base stations in the Tokyo Metropolitan Area for nearly one year.