M3GPSpectra: A novel approach integrating variable selection/construction and MLR modeling for quantitative spectral...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Quantitative analysis of material properties relies on spectral analysis and chemometrics.
• Effective prediction models require feature selection/transformation to reduce noise and redundancy in spectral data.
• Existing methods often use single linear/nonlinear operations, risking information loss.

Purpose of the Study:

• To propose M3GPSpectra, a novel strategy for constructing quantitative analysis models in spectroscopy.
• To integrate feature selection, linear/nonlinear transformation, and model construction into a unified framework.
• To improve the accuracy of prediction models in spectral quantitative analysis.

Main Methods:

• Utilizes a genetic programming algorithm for selecting and reconstructing spectral feature variables.
• Employs multivariate linear regression (MLR) to evaluate variable performance.
• Achieves end-to-end parameter learning through iterative optimization.

Main Results:

• M3GPSpectra demonstrated superior performance compared to seven other popular linear and nonlinear methods across six datasets.
• Generated 19 prediction models, outperforming existing techniques.
• The method proved robust, showing no significant sensitivity to training sample size.

Conclusions:

• M3GPSpectra offers a promising, integrated approach for spectral quantitative analysis.
• The unified framework enhances prediction model accuracy and efficiency.
• The method's effectiveness and robustness make it valuable for analytical chemistry applications.