Postsupervised hard c-means classifier

Abstract

Miyamoto et al. derived a hard clustering algorithms by defuzzifying a generalized entropy-based fuzzy c-means in which covariance matrices are introduced as decision variables. We apply the hard c-means (HCM) clustering algorithms to a postsupervised classifier to improve resubstitution error rate by choosing best clustering results from local minima of an objective function. Due to the nature of the prototype based classifier, the error rates can easily be improved by increasing the number of clusters with the cost of computer memory and CPU speed. But, with the HCM classifier, the resubstitution error rate along with the data set compression ratio is improved on several benchmark data sets by using a small number of clusters for each class. © Springer-Verlag Berlin Heidelberg 2006.