Research On Hidden Markov Model Clustering And It's Applications

Hidden Markov Model (HMM), as one particular class of statistical models used in pattern recognition and clustering, have been applied with great success in both cases. They provide a sound statistical framework, and allow for efficient and numerically stable algorithms. They have become a basic and well-understood tool in the applied sciences.One fundamental problem in the application of HMM is to train the model, particularly in the presence of incomplete data. Training a model is important with regard to good-quality of the clustering.Training a model is the process of adjusting the parameters of a HMM. The objective function is to maximize the likelihood of observing a given observation sequence. Due to the complexity of the likelihood landscape, no efficient, general, global optimization procedures are known.The most widely used training procedure, called the Baum-Welch algorithm, belongs to the class of algorithms dealing with missing data problems. Author focuses mainly on how to improve this algorithm to improve the quality of the clustering. The majority of work is summarized here:A novel clustering method based on frequency-sensitivity and HMM is proposed to assure that the empty clusters or clusters having very few points aren't obtained. Experimental results show this algorithm is more effective for gene expression datasets.Based on both self-splitting and merging competitive learning, author present a new clustering algorithm with HMM. In self-split step, Shannon entropy is used to make each cluster contain more data from one natural cluster. In merge step, clusters which contain the data from the same natural cluster are merged. The new algorithm can improve the quality of the clustering.