Finding Behavior Patterns from Temporal Data using Hidden Markov Model based Unsupervised Classification

Abstract

This paper describes a clustering methodology for temporal data using hidden Markov model(HMM) representation. The proposed method improves upon existing HMM based clustering methods in two ways: (i) it enables HMMs to dynamically change its model structure to obtain a better fit model for data during clustering process, and (ii) it provides objective criterion function to automatically select the optimal clustering partition. The algorithm is presented in terms of four nested levels of searches: (i) the search for the optimal number of clusters in a partition, (ii) the search for the optimal partition structure, (iii) the search for the optimal HMM structure for each cluster, and (iv) the search for the optimal parameter values for each HMM. Preliminary experiments with artificially generated data demonstrate the effectiveness of the proposed methodology.