Post-supervised Fuzzy c-Means Classifier with Hard Clustering
H. Ichihashi, Katsuhiro Honda, 奈歩 桑本, Takao Hattori · 2007
A fuzzy c-means classifier (FCMC) based on a generalized fuzzy c-means clustering with iteratively reweighted least square technique (IRLS-FCM) has been proposed. In this paper, we derive a generalized hard c-means (HCM-g) clustering algorithm by defuzzifying IRLS-FCM. Many hard clustering results are obtained from local minima of the HCM-g objective function. Although HCM-g is not a fuzzy clustering algorithm, it is applied to a fuzzy classifier and the best values of the parameters such as the fuzzifiers were chosen by using golden section search method. Whereas the goal of FCMC is to minimize classification error rate on unseen new test data, the proposed classifier aims at minimizing resubstitution error rate by using only a small number of clusters. The proposed classifier with two clusters for each class achieves low resubstitution error rate on several benchmark data sets