Supervised Change-Point Detection with Dimension Reduction

This paper presents an automated approach for calibrating change point detection algorithms for high-dimensional time series. Our method leverages partial annotations provided by experts to learn a diagonal Mahalanobis metric, combined with a detection algorithm to replicate the expert's segmentation strategy on new signals. Our approach includes sparsity-inducing regularization to improve accuracy, which performs dimension selection and adapts to partial annotations. Our experiments on audio signals and physiological time series signals demonstrate that supervised learning improves detection accuracy significantly.