B1147
Title: Non-parametric distribution-free CUSUM for online change-point detection
Authors: Yao Xie - Georgia Institute of Technology (United States) [presenting]
Haoyun Wang - Georgia Institute of Technology (United States)
Abstract: In modern applications, it is of interest to detect change without making distributional assumptions for using possibly high-dimensional time series due to the complex nature of the data. A general framework of non-parametric CUSUM procedure is developed based on popular distribution-free statistical divergences that can be conveniently estimated by mini-batches of samples, such as MMD and classification loss, computed from mini-batches of data. A way to analyze the statistical performance of such procedures is presented by extending the classic non-linear renewal theory.