Using variable combination population analysis for variable selection in multivariate calibration.

Area of Science:

• Chemometrics
• Data Science
• Machine Learning

Background:

• High-dimensional datasets with limited samples pose challenges for variable selection.
• Existing methods may not efficiently identify the most informative variables.

Purpose of the Study:

• To introduce a novel variable selection strategy: Variable Combination Population Analysis (VCPA).
• To evaluate VCPA's performance against established variable selection techniques.

Main Methods:

• VCPA utilizes an exponentially decreasing function (EDF) to progressively reduce the variable space.
• Binary Matrix Sampling (BMS) generates diverse variable subsets for sub-model construction.
• Model Population Analysis (MPA) identifies optimal subsets based on root mean squares error of cross-validation (RMSECV) and variable frequency.

Main Results:

• VCPA demonstrated strong performance in variable selection across three Near-Infrared (NIR) datasets.
• The strategy effectively determined variable importance based on frequency in top-performing sub-models.
• VCPA showed comparable or superior results when benchmarked against Genetic Algorithm-Partial Least Squares (GA-PLS), Monte Carlo Uninformative Variable Elimination by PLS (MC-UVE-PLS), Competitive Adaptive Reweighted Sampling (CARS), and Iteratively Retains Informative Variables (IRIV).

Conclusions:

• VCPA is a robust and effective strategy for variable selection in high-dimensional chemometric datasets.
• The method offers a valuable alternative for researchers seeking to improve data analysis efficiency and accuracy.
• MATLAB source code for VCPA is available for academic use.