A0627
Title: A general framework for online change point detection in nonparametric regression
Authors: Carlos Misael Madrid Padilla - Washington University in St Louis (United States) [presenting]
Abstract: The aim is to present a general framework for online change point detection in nonparametric regression models. In this setting, data arrive sequentially as covariate response pairs, and the underlying regression function is allowed to change at an unknown point in time. At each time step, the procedure fits two estimators, one to past data and one to recent data, by minimizing empirical squared loss over a chosen function class, and raises an alarm when the discrepancy between these estimators exceeds a time-dependent threshold. This approach accommodates a broad range of function classes, including kernel smoothers, spline-based models, trend filtering methods, and deep neural networks, allowing the procedure to adapt to various structural assumptions and complexities. General theoretical guarantees are established for the proposed method, ensuring control of the false alarm probability and sharp bounds on the detection delay.
