Strategies of handling different variables reduction for LDA

This paper discusses the strategy of conducting variable reduction processes such that they contribute to optimise the performance of linear discriminant analysis (LDA). The variables selection technique with local searching algorithm is manipulated. The technique is proposed to choose useful variables that give minimum error rate on LDA. Meanwhile, principal component analysis is used to extract important information from the original variables. The behaviour of eigenvalue and total variation explained is studied to understand how these two indicators may give optimum performance of LDA. Performance of the proposed strategy and LDA with all variables was assessed in leave-one-out fashion to avoid biasness. This study discovers that LDA with backward elimination is competitive to the full model, but extra concern needs to be given to the PCA