Boosting model performance and interpretation by entangling preprocessing selection and variable selection.

Area of Science:

• Chemometrics
• Data Science
• Analytical Chemistry

Background:

• Data preprocessing is essential for removing artifacts and enhancing information in analytical data.
• Selecting optimal preprocessing strategies is challenging due to numerous available methods.
• Existing methods often focus solely on predictive performance, neglecting model interpretability.

Purpose of the Study:

• To integrate variable selection into a Design of Experiments (DoE)-based preprocessing selection approach.
• To enhance the interpretability and predictive performance of chemometric models.
• To address the impact of data artifacts on variable selection and model interpretation.

Main Methods:

• Utilized a Design of Experiments (DoE) framework for selecting preprocessing strategies.
• Integrated variable selection methods within the DoE-based preprocessing selection process.
• Employed Partial Least Squares (PLS) as the modeling technique and PPRV-FCAM for variable selection.

Main Results:

• The integrated approach significantly improved both model interpretability and predictive performance.
• Variable selection results showed higher compliance with true informative variables compared to individual optimization.
• Demonstrated that combining preprocessing and variable selection yields superior outcomes.

Conclusions:

• The presented generic approach enhances objective interpretation and predictive accuracy in chemometric modeling.
• This integrated strategy offers a robust solution for selecting optimal preprocessing and variable selection methods.
• The methodology is adaptable to various models, variable selection techniques, and preprocessing methods.