High-Dimensional Online Change Point Detection with Adaptive Thresholding and Interpretability

ICLR 2026 Conference Submission18229 Authors

Keywords: Interpretability, Explainability, High-dimensional data, Contrastive Explanations

TL;DR: We propose SWCPD, a robust and interpretable online change point detection framework that leverages random projections, enabling adaptive thresholding, and contrastive explanations for high-dimensional data streams.

Abstract: Change point detection (CPD) identifies abrupt and significant changes in sequential data, with applications in human activity recognition, financial markets, cybersecurity, manufacturing, and autonomous systems. While traditional methods often struggle with the computational demands of high-dimensional data, they also fail to provide explanations for detected change points, limiting their practical usability. This paper introduces a CPD framework that enhances both interpretability and scalability by leveraging the Sliced Wasserstein (SW) distance. Our contributions are fourfold: (1) we present a method to transform multivariate data into one-dimensional time series using the SW distance, enabling compatibility with existing CPD methods; (2) we derive theoretical insights, demonstrating that random slices of the SW distance follow a Gamma distribution, which facilitates statistical hypothesis testing for CPD; (3) we propose a novel self-adapting online CPD algorithm based on an adaptive threshold for a given significance level $\alpha$; and (4) we propose a model-specific framework for generating contrastive explanations for annotated change points. We find that our method outperforms popular (online/offline) change point detection methods by reducing the number of false positives by at least 63% while also providing interpretable change points and maintaining competitive or superior detection performance, making it practical for deployment in high-stakes applications.

Primary Area: interpretability and explainable AI

Submission Number: 18229