Improved variable reduction in partial least squares modelling based on predictive-property-ranked variables and...

Area of Science:

• Chemometrics
• Spectroscopy
• Data Analysis

Background:

• Partial Least Squares (PLS) calibration performance can be enhanced by eliminating uninformative variables.
• Existing Stepwise Variable Reduction methods using Predictive-Property-Ranked Variables (SVR-PPRV) often maintain constant PLS model complexity during reduction.

Purpose of the Study:

• To introduce and evaluate three novel SVR-PPRV methods with adaptable PLS model complexity (PPRVR-CAM).
• To investigate the selective and predictive abilities of these new methods compared to existing techniques.

Main Methods:

• Developed three Predictive-Property-Ranked Variable Reduction with Complexity Adapted Models (PPRVR-CAM) methods.
• Utilized absolute PLS regression coefficients as the predictive property for variable ranking.
• Compared PPRVR-CAM methods against modified SVR-PPRV and reference methods (UVE-GA-PLS, UVE-iPLS) using NIR and simulated datasets.

Main Results:

• The three new PPRVR-CAM methods retained significantly fewer informative variables than existing methods without compromising prediction ability.
• PPRVR-CAM methods demonstrated consistent variable retention, unlike UVE-GA-PLS and UVE-iPLS.
• Variable ranking renewal and complexity adaptation proved beneficial for selective and predictive performance.

Conclusions:

• The newly developed PPRVR-CAM methods offer superior variable reduction performance for PLS calibration.
• Adaptive complexity in variable reduction is advantageous, leading to more efficient models.
• A preferred PPRVR-CAM method is identified for practical application.