A New Feature Transformation Method Based On Genetic Algorithm

In pattern recognition, feature transformation methods transform features to a new space with aim of obtaining more discriminative or orthogonal features and so more separable classes. In this paper, a linear feature transformation method is presented which considers both of class discrimination and features orthogonality criteria, simultaneously. This feature transformation is obtained using genetic algorithm where its fitness function is determined using mentioned criteria. For feature discrimination criterion, Dunn index is used. On the other hand, for feature orthogonality and independency, covariance matrix and ratio of sum of its diagonal elements to sum of non-diagonal elements is used. We use these criteria to determine the genetic algorithm fitness function. Experiments on UCI dataset show that the proposed feature transformation performs better than or as well as other known linear transformation methods.