Dimensionality Reduction with Neuro-Fuzzy Discriminant Analysis

One of the most important tasks in any pattern recognition system is to find an informative, yet small, subset of features with enhanced discriminatory power. In this paper, a new neuro-fuzzy discriminant analysis based feature projection technique is presented based on a two stages hybrid of Neural Networks, optimized with Differential Evolution (DE), and a proposed Fuzzy Linear Discriminant Analysis (FLDA) technique. Although dimensionality reduction via FLDA can present a set of well clustered features in the reduced space, but like any version of the existing DA’s it assumes that the original data set is linearly separable, which is not the case with many real world problems. In order to overcome this problem, the first stage of the proposed technique maps the initially extracted features in a nonlinear manner into a new domain, with larger dimensionality, in which the features are linearly separable. FLDA acts then on these linearly separable features to further reduce the dimensionality. The proposed combination, referred to as NFDA, is validated on a prosthetic device control problem with Electroencephalogram (EEG) datasets collected from 5 subjects achieving a maximum testing accuracy of 85.7% for a three classes of EEG based imaginations of movements. Keywords—Feature Projection, Linear Discriminant Analysis, Prostheses Control.