Segmenting Time Series Data Using GP-HSMM with Nonparametric Bayesian Model

Abstract

In this paper, we propose a method for dividing continuous time-series data into segments in unsupervised manner. Humans recognize perceived continuous information by dividing it into signicant segments such as words and unit motions. To this end, we have been proposed a method based on hidden semi-Markov model with Gaussian process (GP-HSMM). However, it has a big drawback that it requires the number of classes into which time-series data is segmented. To overcome this problem, in this paper, we extend GP-HSMM to nonparametric Bayesian model by introducing hierarchical Dirichlet processes (HDP), and propose hierarchical Dirichlet processes-Gaussian process-hidden semi-Markov model (HDP-GP-HSMM). Hence, the infinite number of classes is assumed and the number of classes are estimated by applying slice sampling. In the experiment, we used the various time-series data and showed that our proposed model can estimate more correct number of classes and achieve more accurate segmentation than baseline methods.