Reevaluating feature importance in machine learning models for CO2 photoreduction: A statistical perspective

Chen et al. have advanced the theoretical design of dual-site metallo-covalent organic frameworks for enhancing CO2 photoreduction into C2H4 using various machine learning algorithms. While they demonstrated high predictive accuracy using a stacking approach with seven selected algorithms, this study emphasizes the potential biases in feature importance derived from these models. It argues for the necessity of computing unbiased feature importances and highlights the complications posed by different methodologies across models. Further, it recommends robust statistical techniques, such as Spearman's correlation and Kendall's tau, to improve interpretability and validity. Addressing collinearity through Variance Inflation Factor (VIF) analysis is also crucial. These steps aim to deepen understanding and optimize machine learning applications for carbon capture and utilization.