A1111
Title: Random forests for change point detection
Authors: Malte Londschien - ETH Zürich (Switzerland) [presenting]
Peter Buehlmann - ETH Zurich (Switzerland)
Solt Kovacs - ETH Zurich (Switzerland)
Abstract: Changeforest, a novel nonparametric multivariate change point detection method, is introduced. Change point detection considers the localization of abrupt distributional changes in time series. This has bioinformatics, neuroscience, biochemistry, climatology, and finance applications. The power of modern nonparametric classifiers like random forests is leveraged by reframing the change point detection problem as a supervised learning task. A log-likelihood ratio that uses random forests' class probability predictions is constructed to compare change point configurations and pair this with a computationally feasible search method. It is proved that Changeforest consistently locates change points in single change point scenarios. In a comprehensive simulation study, changeforest achieves improved empirical performance compared to existing methods.
