Bhattacharyya distance feature selection

Abstract

A recursive algorithm called Bhattacharyya distance feature selection for selecting a real-optimum feature under normal multidistribution is presented. The key of this method is to change the problem of minimizing the criterion of the sum of the upper bound of error probability of every two class pairs in subspace to a problem of solving a nonlinear matrix equation in a multiclass problem under an orthonormal coordinate system. The recursive algorithm is considered as finding the optimal solution of a transformation matrix from the nonlinear matrix equation. The theoretical analysis and experimental results show that under normal multidistribution the performance of the proposed algorithm is superior to the performance of any previous one.