An improved kernel-induced possibilistic fuzzy c-means clustering algorithm based on dispersion control

Abstract

Presented is a fuzzy clustering algorithm based on adaptive kernel methods. To utilize benefits of combining fuzzy c-means (FCM) and possibilistic c-means (PCM) models, we adopt the possibilistic fuzzy c-means (PFCM) model that produces memberships and possibilities simultaneously for each cluster while clustering unlabeled data. As an extension of kernel-induced PFCM (KPFCM), we propose an improved kernel-induced possibilistic fuzzy c-means (IKPFCM) algorithm. With the kernel methods, the input space can be implicitly mapped into a high-dimensional feature space in which the nonlinear patterns appear linear. The main feature of kernel induced models, compared to other fuzzy clustering models such as FCM, PCM and PFCM using Euclidean distance, is that they are based on Gaussian kernel-induced non-Euclidean distance. For ameliorating the performance of KPFCM, IKPFCM uses the approach that the Gaussian width parameter is selected randomly in a suitable range at each iteration. The experimental results show that the proposed IKPFCM algorithm achieved significantly better or sometimes similar clustering performance than its competitors considered.