Title:Changepoint detection for high-dimensional time series with missing data

Abstract:This paper describes a novel approach to changepoint detection when the observed high-dimensional data may have missing elements. The performance of classical methods for changepoint detection typically scales poorly with the dimensionality of the data, so that a large number of observations are collected after the true changepoint before it can be reliably detected. Furthermore, missing components in the observed data handicap conventional approaches. The proposed method addresses these challenges by modeling the dynamic distribution underlying the data as lying close to a time-varying low-dimensional submanifold embedded within the ambient observation space. Specifically, streaming data is used to track a submanifold approximation, measure deviations from this approximation, and calculate a series of statistics of the deviations for determining when the underlying manifold has changed in a sharp or unexpected manner. The approach described in this paper leverages several recent results in the field of high-dimensional data analysis, including subspace tracking with missing data, multiscale analysis techniques for point clouds, online optimization, and changepoint detection performance analysis. Simulations and experiments highlight the robustness and efficacy of the proposed approach in detecting an abrupt change in an otherwise slowly-varying low-dimensional manifold.