Model-Based Clustering of Time Series Exhibiting Nonlinear Dynamics

Abstract

We consider the problem of time series clustering, which asks how to cluster objects with dynamically changing properties over time into sensible groups in an unsupervised manner. In tackling this problem, we pursue a model-based approach, introducing a mixture of nonlinear state-space models and corresponding inference algorithms. In particular, we develop a Monte Carlo expectation-maximization algorithm for finite mixtures and a Markov chain Monte Carlo algorithm for both finite and infinite mixtures. We apply our work to a variety of examples in computational neuroscience and demonstrate the utility of time series clustering in these real-world settings.