Hierarchical Regime Switching Dynamic Matrix Factor Models for Modeling Mouse Motion Behavior

• Activity Number: 42 - Analysis of Dynamic High-Dimensional Data
• Type: Invited
• Date/Time: Sunday, August 8, 2021 : 3:30 PM to 5:20 PM
• Sponsor: IMS
• Author(s): Rong Chen*
• Companies: Rutgers University
• Keywords: motion image; regime switching; Bayesian; factor model; matrix time series

Abstract

In this talk we present a dynamic matrix factor model with a regime switching factor process under a hierarchical regime structure. It is designed to model and analyze sequences of aligned video images of mouse moving in a bucket at 30ms per frame. The images are presented by 80 by 80 matrices. A matrix factor model is applied for reducing the observed image sequence to a smaller matrix factor process. The factor process is then modelled with a regime switching matrix AR model, where the regimes are controlled by a three-layer motion decision structure: basic motion module, motion motif, and action. Each basic module (similar to the alphabets in text) is uniquely defined by the parameter set in the Matrix AR model, a motion motif (similar to words) consists of a small number of modules with a Markovian transition matrix, and action (similar to a sentence) is a sequence of motifs. Based on the observed images, we will identify and estimate the models under a Bayesian framework. It is used to detect abnormal motion behavior of diseased mouse.