Feature reduction based on analysis of fuzzy regions

In this paper a novel approach to feature reduction is proposed, based on analysis of class regions generated by a fuzzy classifier. It is shown how the degree of overlaps in the class regions, or the degree of exceptions inside the fuzzy rules generated by the fuzzy classifier, is used for feature evaluation. To measure such degrees, an exception ratio is defined. Given a set of features, a subset of features that has the lowest sum of the exception ratios has a tendency to contain the most relevant features, compared to other subsets with the same number of features. The proposed algorithm eliminates irrelevant features. Given a set of remaining features, the proposed algorithm eliminates the next feature, the elimination of which minimizes the sum of the exception ratios. Experiments show that the proposed algorithm effectively eliminates irrelevant features; its performance compares favorably with that of a previous algorithm