Representative subset selection in modified iterative predictor weighting (mIPW) — PLS models for parsimonious multivariate calibration

A new sample selection technique, named as subset selection based on weighted X (instrumental responses) and y (concentrations) joint distances (SSWD), is proposed in this work. The SSWD takes into account exact contributions of both X and y spaces in calibration, thus being sensitive to variations of both instrumental signals and concentrations. In order to avoid the interference of redundant wavelengths, uninformative wavelengths are eliminated at the first stage by using the modified iterative predictors weighting PLS (mIPW-PLS), then, the SSWD is performed based on the mIPW selected wavelengths to extract the optimal calibration subset. There is no subjective parameter that should be adjusted in calculation of the mIPW-SSWD-PLS, which makes it convenient to perform in practice. To validate the effectiveness and universality of the strategy, it was applied to two different sets of Near-infrared (NIR) spectra. The results indicated that the final model was constructed economically because only the most informative wavelengths and samples were used. The study reveals that the proposed method is of value to reduce both the interference and complexity of multivariate calibration involving large and complex matrices.