Non-Parametric Online Changepoint Detection Algorithm

Changepoint detection using CUSUM is a well-studied problem and some useful variations and enhancements to the basic CUSUM algorithm have been proposed. However, CUSUM algorithms require the user to specify parameter(s) specific to different signals. This paper presents a set of additions to the CUSUM algorithm that eliminates sensitivity parameters using several new automation techniques, allows easy extensibility to generic (i.e. more than just mean) changepoint detection, and makes it more adaptive, while maintaining causality at all times. We automated the estimation of the CUSUM threshold, h, using a Markov model that updates h at each time step to reflect the most up-to-date information. The expected change parameter, delta, is estimated using a regression method. We perform experiments to investigate segmentation capabilities and compare performance with that of Bayesian Online Changepoint Detection.