A Backward Variable Selection method for PLS regression (BVSPLS)

Abstract

Variable selection has been discussed in many papers and it became an important topic in areas as chemometrics and science in general. Here a backward iterative step-by-step wrapper method is proposed using PLS. The root-mean-square error of prediction (RMSEP) for an independent test set is used as selection criterion to quantify the gain obtained using the selected range of variables. The method has been applied to different data sets and the results obtained revealed that one can improve or at least keep constant the prediction performances of the PLS models compared to the full-spectrum models. Moreover with the advantage that the number of variables is reduced driving to an easier interpretation of the relationship between model and sample composition and/or properties. The aim is not to compare to other variable selection methods but to show that a simple one can improve or at least keep constant the prediction performances of the PLS models by using only a limited number of variables.