Improved variable reduction in partial least squares modelling by Global-Minimum Error Uninformative-Variable Elimination | JoVE Visualize

Area of Science:

Chemometrics
Machine Learning
Data Analysis

Background:

Partial Least Squares (PLS) regression is widely used for calibration.
Variable selection methods like Uninformative-Variable Elimination for PLS (UVE-PLS) aim to improve PLS performance by removing irrelevant variables.
Classical UVE-PLS often retains a large number of variables, limiting model parsimony.

Purpose of the Study:

To introduce and evaluate a modified Uninformative-Variable Elimination method, Global-Minimum Error Uninformative-Variable Elimination for PLS (GME-UVE-PLS).
To achieve greater variable reduction and potentially enhance the predictive ability of PLS models compared to existing methods.
To obtain more parsimonious and selective PLS models.

Main Methods:

The proposed GME-UVE-PLS method involves repeating Uninformative-Variable Elimination until no further variable reduction is possible, followed by a search for the global minimum of the root mean squared error of cross-validation (RMSECV).
The predictive ability of PLS models is assessed using RMSECV after each iteration.
Performance was evaluated on simulated, NIR, NMR, and molecular descriptor datasets, resulting in twelve profile-response calibrations.
Statistical comparisons were made against UVE-PLS and 1-step UVE using one-sided paired t-tests.

Main Results:

GME-UVE-PLS typically eliminates significantly more variables than classical UVE-PLS.
The predictive abilities of PLS models built with GME-UVE-PLS selected variables were generally superior.
The method successfully identified smaller, more parsimonious variable sets with comparable or improved predictive performance.
Fewer uninformative variables, lacking chemical meaning for the response, were retained compared to UVE-PLS.

Conclusions:

GME-UVE-PLS offers improved selectivity and parsimony over classical UVE-PLS for Partial Least Squares regression.
The method enhances the efficiency of variable reduction in PLS calibration.
GME-UVE-PLS provides a valuable alternative for developing more robust and interpretable PLS models.