Plug-and-Play Bayesian Online Change-Point Detection in Gaussian Processes
Keywords: Bayesian decision-making under uncertainty, Non-stationarity in online time-series.
Abstract: We study the problem of detecting non-stationarity in online time series, when the underlying distribution is assumed to be a piecewise 1d-Gaussian process. Drawing inspiration from Bayesian online change-point detection methods such as that of [Adams and MacKay, 2007], we construct a restarted variant of that to specifically deal with arbitrary changes both in mean and variance of 1d-Gaussian processes. We evaluate our algorithm on both synthetic datasets of varying task difficulty and on prevalent real-world data across a variety of fields. Our results compare favorably with state-of-the-art, as measured by the detections' F1-Score. Code will be provided to ensure easy reproducibility.
Submission Number: 106
Loading