Comparative Performance of Classical Fisher Linear Discriminant Analysis and Robust Fisher Linear Discriminant Analysis

Linear discriminant analysis for multiple groups can be performed using Fisher's technique which can be applied to classify and predict observations into various populations. Classical Fisher linear discriminant analysis (FLDA) is highly susceptible to outliers. The poor performance of classical FLDA is due to lack of robustness of the classical estimators used to train the model. The proposed robust FLDA combine the features of classical FLDA and weighted sample observations. This paper examines the comparative classification performance of Fisher linear discriminant analysis and the proposed robust Fisher linear discriminant analysis. The paper focuses on the influence scaled normal and unscaled normal data set have on the classical Fisher and the robust Fisher techniques. The objectives of this paper are to compare the classification performance of these methods based on the mean of correct classification and to examine the separation between the group means. The classification results indicate that the proposed procedure has improved classification rate compared to the classical Fisher linear classification analysis. The simulation showed that both procedures have comparable separation capability.