A Subset Selection Algorithm for Multivariate Modeling Based on the Spectral Variations

This paper proposes a subset selection method, named sample set partitioning based on joint x-y-z distances (SPXYZ) algorithm, for multivariate modelling. The proposed method is a modified version of the original sample set partitioning based on joint x-y distances (SPXY) algorithm. The contributions from the dependent variable (z) space for parameters that cause the systematic error in measured spectra, including external factors and inherent characteristics, are added to the original SPXY algorithm. Here, the z differences denotes the variability in the dimension of external disturbances and inherent characteristics. Based on two real world datasets, SPXYZ is employed with partial least-squares (PLS) to demonstrate the advantages of subset selection by adding the contributions from external factor, i.e., temperature and inherent characteristic, i.e., background components. We compare the prediction performance of SPXYZ-PLS model with other three PLS models using random sampling (RS), Kennard-Stone (KS) and SPXY. The prediction performance results from experimental studies showed that the prediction performance of SPXYZ-PLS is significantly better than the other models. Therefore, the proposed method is an alternative method of subset selection for calibration modeling.