Model-Based Time Series Clustering Using CHOMP

• Activity Number: 241
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Yuan Zhuang*+ and Nicole Lazar
• Companies: University of Georgia and University of Georgia
• Keywords: model-based ; cluster analysis ; time series ; higher order ; markov process ; non-linear

Abstract

Time series clustering has attracted lots of interest in recent years due to its ability to cluster temporal patterns in data. Among all approaches, our focus is on model-based algorithms since it can study the statistical properties which is usually unknown in other clustering algorithms. Previous work has used models such as ARIMA, simple Markov Chain, and Hidden Markov Model (HMM). However, these models can only account for simple dependence structure in the time series. This may result in errors in the clustering stage since the complicated data structure is not characterized appropriately by the model. In this paper, we propose a model-based time series clustering using the Copula-based Higher Order Markov Process (CHOMP) which allows for higher order non-linear dependence structure and accepts continuous random variables. The algorithm combines the finite mixture model approach and iterative relocation method to obtain the most probable clustering based on the likelihood. We further relax the stationary condition in the CHOMP so that it can be applied to some non-stationary time series. Results from a simulation study using the proposed method will be presented in the talk.