An Ensemble of Variable Influence on Projection of Pls for Wavelength Selection

A new strategy named the ensemble of variable projection importance by partial least squares regression (PLS-EVIP) was proposed to select the variables from multivariate calibration of spectral analysis. The proposed PLS-EVIP extracted the most informative variables by introducing a fusion of ensemble strategy and the threshold setting to the selected variables. To demonstrate the effectiveness of the proposed PLS-EVIP strategy, a sets of experimental spectra datasets were applied to train the PLS-EVIP. The results demonstrated that the performance model could be significantly improved by the proposed PLS-EVIP strategy. It was also revealed that the model ensemble approach provided a robust collection way for the information from individual model. Compared with the other method, PLS-EVIP could keep a certain number of the more informative variables for the final modeling and the gave satisfactory results. It is expected that ensemble strategy combined with the threshold setting approach are promising methods in chemometrics and are developed for novel implementations in other studies.