Reevaluating feature importance in gas–solid interaction predictions: A call for robust statistical methods

Abstract

Liu et al. conducted an insightful investigation into feature importance analysis for predicting CH4 adsorption isotherms in metal–organic frameworks (MOFs), revealing key geometric features that influence model predictions. While their use of advanced machine learning techniques, including neural networks and extra tree regression (ETR), achieved notable accuracy, concerns arise regarding the model-specific biases in feature importance metrics. This paper critically evaluates these metrics, highlighting the risks of misinterpretation due to the lack of ground truth validation. We advocate for the adoption of bias-free statistical methods, such as Spearman's rank correlation and Kendall's tau, which offer a more reliable framework for assessing feature importance. Implementing these approaches could enhance the understanding of gas–solid interactions and improve the reliability of machine learning applications in this domain.