Variable selection by modified IPW (iterative predictor weighting)-PLS (partial least squares) in continuous wavelet regression models

Variable selection is often used to produce more robust and parsimonious regression models. But when they are applied directly to the raw near-infrared spectra, it is not easy to select appropriate variables because background and noise will often overshadow or overlap the absorption bands of analyte. In this work, a new hybrid algorithm based on the selection of the most informative variables in the continuous wavelet transform (CWT) domain is described. The strategy is a combination of CWT and a procedure of modified iterative predictor weighting-partial least square (mIPW-PLS). After elimination of the background and noise in NIR spectra by CWT, the mIPW-PLS approach is used to select the most informative CWT coefficients. With the selected CWT coefficients, a PLS model is built finally for prediction. It is indicated that the extraction of most important variables in the CWT domain can effectively avoid the interference of background and noise, and result in a high quality of regression model with a very small number of variables and fewer PLS components.