Density based Fuzzy Support Vector Machines for Multicategory Pattern Classification

Abstract

Support vector machines (SVMs) are known to be useful for separating data into two classes. However, for the multiclass case where pairwise SVMs are incorporated, unclassifiable regions can exist. To solve this problem, Fuzzy support vector machines (FSVMs) was proposed, where membership values are assigned according to the distance between patterns and the hyperplanes obtained by the “crisp” SVM. However, they still may not give proper decision boundaries for arbitrary distributed data sets. In this paper, a density based fuzzy support vector machine (DFSVM) is proposed, which incorporates the data distribution in addition to using the memberships in FSVM. As a result, our proposed algorithm may give more appropriate decision boundaries than FSVM. To validate our proposed algorithm, we show experimental results for several data sets.