Canonical sequence extraction and HMM model building based on hierarchical clustering 1

This work presents an HMM based hierarchical clustering method, aiming at the extraction of typical temporal sequences, and outliers are discarded in the process of clustering. After the Baum-Welch training step of HMM, TWM (transition weighted matrix) is used as the features of sample sequences, thus the original clustering problem is converted to a relatively easy problem of points clustering in a high dimensional space. By using hierarchical clustering and NCut (normalized cut) method, the unsteadiness in separation is efficiently prevented and the time consuming is relatively small. The method is used in unsupervised learning of typical hand gestures and facial expressions.