An Interweaved HMM/DTW Approach to Robust Time Series Clustering

Abstract

We introduce an approach for model-based sequence clustering that addresses several drawbacks of existing algorithms. The approach uses a combination of Hidden Markov Models (HMMs) for sequence estimation and Dynamic Time Warping (DTW) for hierarchical clustering, with interlocking steps of model selection, estimation and sequence grouping. We demonstrate experimentally that the algorithm can effectively handle sequences of widely varying lengths, unbalanced cluster sizes, as well as outliers.