A Backward Variable Selection method for PLS regression (BVSPLS).

Area of Science:

• Chemometrics
• Analytical Chemistry
• Data Science

Background:

• Variable selection is crucial in chemometrics and scientific modeling.
• Existing methods can be complex; simpler approaches are needed for model interpretability.

Purpose of the Study:

• To propose a backward iterative step-by-step wrapper method for variable selection using Partial Least Squares (PLS).
• To demonstrate that this method can maintain or enhance prediction performance compared to full-spectrum models.
• To highlight the benefit of reduced variable sets for improved model interpretability.

Main Methods:

• A backward iterative wrapper method was developed using PLS.
• The root-mean-square error of prediction (RMSEP) on an independent test set served as the selection criterion.
• The method was applied to various datasets to evaluate its effectiveness.

Main Results:

• The proposed variable selection method improved or maintained the prediction performance of PLS models.
• Models built with selected variables showed comparable or better predictive accuracy than full-spectrum models.
• A significant reduction in the number of variables was achieved, enhancing model interpretability.

Conclusions:

• A simple backward iterative variable selection method using PLS is effective.
• This method offers a practical approach to enhance PLS model interpretability and maintain predictive power.
• The technique provides a valuable tool for variable selection in chemometrics and related fields.