Data driven feature extraction based on parameterized transformations of representation space
Abstract
To analyze a stochastic process described by samples drawn from different classes a method for automatic extraction of discriminant features in reduced dimension space is proposed. To be effective dimension reduction should be achieved with minimum loss of information. The proposed method is based on the search for an optimal transformation between representation space and feature space according to class information. Information is measured using a mutual information estimate. A nonparametric entropy estimate and a stochastic distributed optimization algorithm are used to solve this problem. An experimental study of a classification problem of specific waveforms in sleep EEG assesses the efficiency of the proposed method.