Agrupamento fuzzy baseado no kernel de mahalanobis com distâncias quadráticas adaptativas

Abstract

In this work, we propose fuzzy clustering methods based on a Mahalanobis-type kernel with adaptive quadratic distances de ned either by fuzzy covariance matrices that are commom to all clusters or by a fuzzy covariance matrix that can be different for each cluster. These matrices can be also full or diagonal. The Mahalanobis-type kernel functions were built from adaptive quadratic distances parameterized by symetric, positive-defined matrices that changes at each iteration of the algorithm. The proposed algorithms were evaluated and compared with the main clustering methods in the literature as well as with their kernelized versions through experiments with simulated and real data.