Abstract: In this paper, we propose a model-based, competitive learning procedure for the clustering of variable-length sequences. Hidden Markov models (HMMs) are used as representations for the cluster centers, and rival penalized competitive learning (RPCL), originally developed for domains with static, fixed-dimensional features, is extended. State merging operations are also incorporated to favor the discovery of smaller HMMs. Simulation results show that our extended version of RPCL can produce a more accurate cluster structure than $k$-means clustering.
Proceedings of the International Conference on Pattern Recognition (ICPR), vol 2, pp.195-198, Barcelona, Spain, September 2000.
Postscript: http://www.cs.ust.hk/~jamesk/papers/icpr00.ps.gz