Random Forests for Change Point Detection
Authors: Malte Londschien, Peter Bühlmann, Solt Kovács; Published in 2023.
Abstract
This study introduces a new method for detecting multiple change points using classifiers in a multivariate nonparametric setting. The method utilizes a classifier log-likelihood ratio that compares different change point configurations based on class probability predictions. The proposed method, called changeforest, is specifically designed for random forests but can also be used with other classifiers that provide class probability predictions, such as the k-nearest neighbor classifier. The study proves the consistency of the method in locating change points in single change point scenarios when paired with a consistent classifier. Extensive simulation studies demonstrate that the proposed method, changeforest, outperforms existing multivariate nonparametric change point detection methods. The changeforest software package provides an efficient implementation of the method for R, Python, and Rust users.
[abs]