Feature selection: a neuro-fuzzy approach

This article attempts to integrate the merits of fuzzy set theory and artificial neural networks under the heading “neuro-fuzzy computing”. The paper describes a method of ranking the features (or subsets of features) using a new fuzzy set theoretic feature evaluation index, and its performance with an existing one is compared. It then presents a neuro-fuzzy approach where a new connectionist model has been designed in order to optimize the fuzzy evaluation index described, which incorporates weighted distance for computing class membership values. This optimization process results in a set of weighting coefficients representing the importance of the individual features. These weighting coefficients lead to a transformation of the feature space for better modeling the class structures. The effectiveness of the algorithm is demonstrated on a speech recognition problem