Change Point Detection for Time Series Data in Complex Systems

Abstract

In this work, we present a change point detection (CPD) method to detect abrupt changes in time-series data obtained from complex systems such as large scale networks. The proposed method works by converting the original time-series into binary-valued sequences with Os and 1s and then identifying the time instances that the density of 1s change. Under a mild assumption that the 0/1 samples are drawn from the same distribution in both reference and test period, we develop a double-direction detection method to detect upward and downward change of the density of 1-samples. The proposed CPD method is applied to operate at both fast and slow time scales to detect changes that last for shorter and longer durations. Numerical results obtained from time-series dataset of large scale cellular network are used to evaluate the performance of the proposed method.