A fuzzy clustering based algorithm for feature selection

This paper deals with a wrapper approach to the problem of feature selection for classification. Based on fuzzy clustering, we develop a new algorithm that operates by testing the error between the cluster structure of the subspace data set and the class structure of the original data set. The true number of clusters in the subspace data set introduces accurate cluster structure information. The classification error rate, based on the difference between the number of clusters in the subspace data set and the number of classes in the original data set, provides a fair evaluation of how well the subset of features represents the original feature set. The experimental results show the advantage of our new algorithm.