Clustering Based on Multiple Criteria for LVQ and K-Means Algorithm

Abstract

When classifying linearly separable data by learning vector quantization (LVQ) or K-Means algorithm (KMA), we cannot necessarily obtain satisfactory classification results for bad selections of initial cluster centers and differences among the distributions of class data. In this paper, to realize reliable classification, clustering based on multiple criteria for LVQ and KMA is proposed, and its performance is provided. To obtain suitable cluster centers, KMA with the split and merge procedure proposed by Kaukoranta et al. is introduced to minimize the squared-error distortion. LVQ using those cluster centers as initial ones is applied to the data, andΚclusters are produced. Introducing a criterion of whether each cluster reveals unimodality, subclusters split by KMA for clusters having no unimodality are merged into appropriate neighboring clusters except one subcluster, and the validity of the classification result is checked.