Improving Linear Regression Performance for Interpretable Control

A Partial Dependence Plot Approach

Abstract

This paper addresses the challenge of improving the performance of interpretable linear models for real-time process control in industrial and chemical systems. While black-box models such as Random Forest and XGBoost achieve high predictive accuracy, their lack of interpretability and limited suitability for real-time applications make them difficult to integrate into production environments. To overcome this limitation, we propose a novel feature engineering approach that leverages Partial Dependence Plots (PDP) to capture the complex nonlinear relationships learned by black-box models and convert them into features appropriate for linear regression. We evaluate the proposed methodology using the well-known publicly available wine quality dataset, which provides a strong benchmark due to its multivariate and nonlinear characteristics that are analogous to those observed in industrial processes. Our results show that PDP-based feature transformation substantially improves the predictive accuracy of linear models, achieving an R² score comparable to that of black-box models. This study proposes a practical solution for building high-performance yet interpretable models, showing strong potential for real-time deployment in process control and monitoring.