HIERARCHICAL CLUSTERING USING DETERMINISTIC ANNEALING

This paper presents a new approach to the problem of hierarchical clustering. The method implements an approximation to joint optimization over all levels of the hierarchy, utilizing deterministic annealing to improve the clustering solution. Similar to the splitting algorithm, cluster nodes at all tree levels are placed at generalized "region"centroids. In this method, though, the node centroids are updated to explicitly enforce desired classification at the leaves, and to approximate the unconstrained clustering solution. The approach is demonstrated to avoid local minima that trap the splitting algorithm and to obtain performance improvement for a normal mixture source and a speech source.