Agrupamento subtrativo baseado em Kernel para dados simbólicos da natureza intervalar

Abstract

In this work, we present extensions for known subtractive clustering methods. The subtractive clustering method for symbolic interval data (iSBC) as an extension of the subtractive clustering method developed by Chiu (1994), as well as the kernelbased subtractive clustering methods defined by one or two components for symbolic interval data (iKSBC1C and iKSBC2C, respectively) as extensions of a kernel-based subtractiveclusteringmethodproposedbyKimetal. (2005). Inaddition, sixstrategies will be proposed: the centroids of the proposed methods will be given as inputs to the methods K-means for interval data based on L2 distance proposed by De Carvalho, Brito and Bock (2006) (iKM+iSBC, iKM+iKSBC1C and iKM+iKSBC2C) and kernel K-means for symbolic data of the interval-valued developed by Costa (2011) (iKKM+iSBC, iKKM+iKSBC1C and iKKM+iKSBC2C) as a way to minimize the sensitivity of these methods to the choice of the centroid for de nition of the initial partition. Experiments using real data showed that the proposed kernelbased subtractive clustering methods (iSBC1C and iSBC2C) obtained better performance than the iSBC method, as well as the K-means (iKM+iSBC, iKM+iKSBC1C and iKM+iKSBC2C) and kernel K-means (iKKM+iSBC, iKKM+iKSBC1C and iKKM+iKSBC2C) methods, both for symbolic data interval-valued, using the centroids of methods proposed as inputs for them also obtained better performance that the iKM and iKKM methods.