Iterative shrinking method for generating clustering

Abstract

The pairwise nearest neighbor method (PNN) generates the clustering of a given data set by a sequence of merge steps. We propose an alternative solution for the merge-based approach by introducing an iterative shrinking method. The new method removes the clusters iteratively one by one until the desired number of clusters is reached. Instead of merging two nearby clusters, we remove one cluster by reassigning its data vectors to the neighbor clusters. We retain the local optimization strategy of the PNN by always removing the cluster that increases the cost function least. We give six alternative implementations, which all outperform the PNN in quality.