报告题目：Automatic Change Point Detection and Segment Estimation via Variational Bayesian Model Selection

报告时间：2022-03-31  09:00 - 09:50

报告人：彭衡  香港浸会大学

ZOOMID：993 6445 5975   密码：220331

报告入口：https://zoom.us/j/99364455975?pwd=VUlvalBWcUtvc3M0aU1IbzJYUUdHUT09

Abstract：Change-point detection has long been an active research area, especially in the Big Data era, where data streams are usually non-stationary. However, practical applications present many new challenges, e.g. multi-dimensional time series, no prior knowledge of the number of change-points, and subsequent estimations for within-segment characteristics. In this talk, we introduce the Bayesian Change-Point Model and propose a scalable variational EM algorithm. Our algorithm can automatically do change-point detection, subsequent inference, and hyperparameter tuning. The comprehensive simulations for a normal mean-variance shift model and a discrete Poisson model as well as a real application in finance demonstrate our advantages over existing methods in both estimation accuracy and computational efficiency.