Selection of the simpler calibration model for multivariate analysis by partial least squares

This paper studies the effect of the elimination of both samples and variables in partial least squares (PLS) analysis with the aim of achieving improved prediction ability in multivariate determination by molecular fluorescence spectrometry. The methodology uses the instrumental responses of three training sets of samples. a central composite design (CCD), a Box-Behnken design (BBD) and a randomised design (RD), to select the optimal subset, which are then submitted to chemical analysis and calibration. In all cases, the PLS models with a reduced number of samples provided accurate results. After, variable selection was applied on these reduced models. The feature selection method was based on the loading weights of the PLS models. It is concluded that selection of calibration sample subset and wavelength selection can improve the prediction ability of multivariate calibration models.