A0486
Title: Nonparametric sequential change-point detection in high dimensions
Authors: Shubhadeep Chakraborty - University of Washington, Seattle, USA (United States) [presenting]
Abstract: Sequential change-point detection is a classical problem in statistics where the goal is to detect a change in the data generating mechanism as soon as possible after it occurs. We propose two nonparametric algorithms and stopping rules to sequentially detect general distributional changes along a high-dimensional data stream, rather than detecting changes only in the mean or in the covariance structure. The first algorithm adopts a sliding window-based approach that performs a sequence of two-sample tests for homogeneity between a post-change sample and a reference pool for every newly recorded observation. A theoretical approximation of the average run length is rigorously derived, which enables us to easily obtain the value of the threshold for the stopping rule, thus achieving control on the false alarm rate. We also establish an explicit upper bound for the expected detection delay and illustrate the impact of certain key factors on the expected detection delay. In the second algorithm, we construct a sequence of thresholds for the sequential two-sample tests using generalized alpha investing rules, controlling the false discovery rate. Numerical studies illustrate the superior performance of our algorithms over other state-of-the-art methods.
