Genetic Algorithm-Based Partial Least-Squares with Only the First Component for Model Interpretation.

Area of Science:

• Multidisciplinary applications in molecular, material, process design, and control.
• Development of interpretable predictive modeling techniques.

Background:

• Accurate predictive models are crucial for design and control.
• Multicollinearity in explanatory variables complicates model interpretation, even in linear models.
• Existing methods struggle to balance predictive power with clear interpretability of variable contributions.

Purpose of the Study:

• To propose a novel modeling approach for enhanced interpretability in predictive models.
• To develop a method that allows regression coefficients to accurately represent the contributions of explanatory variables (X) to objective variables (y).
• To introduce a genetic algorithm-based approach for selecting optimal variable combinations for the proposed model.

Main Methods:

• Focus on a simplified partial least-squares model using only the first component (PLSFC).
• Integration of a genetic algorithm (GA) to select explanatory variables (X) for constructing a predictive PLSFC model, termed GA-PLSFC.
• Validation using simulated datasets and real-world material data.

Main Results:

• The PLSFC model enables regression coefficients to be interpreted as direct contributions of X to y.
• GA-PLSFC successfully identifies variable combinations that yield predictive and interpretable models.
• Effectiveness demonstrated through numerical simulations and practical material science applications.

Conclusions:

• The proposed GA-PLSFC method achieves both high predictive ability and high interpretability.
• Regression coefficients in the GA-PLSFC model serve as reliable indicators of variable contributions.
• The developed technique offers a valuable tool for clarifying phenomena and elucidating mechanisms in various design fields.