Bayesian online change point detection with contextual time series modelling of gaussian process regression

Abstract

We study the change point detection of dynamic system. Since this system is delivered as the time series data, we focus the varying characteristic of the data over the time. The existing change point detection of dynamic system mainly approaches this problem by the simple threshold. However, we choose the Bayesian Online change point detection method to consider time-dependent alarm guideline. In this process, we suggest the prediction model called UPM to more precisely predict the two distinct characteristic data in a week. In addition, we suggest the new hazard function to consider the model's changeable reliability depending on time. Through this approach, we intend to take a conservative stance to determine the change point at unreliable prediction time. As a result, we can suggest the more accurate and cautious online change point detection method.