Enhancing feature importance analysis in battery research: a statistical methods perspective on machine learning limitations

Abstract

This paper addresses critical concerns related to feature importance analysis in battery research, specifically examining the limitations of machine learning-derived feature importances as reported by Yuan et al. While recent studies have achieved impressive prediction accuracy in battery modeling, this paper underscores that such accuracy does not necessarily ensure the trustworthy interpretation of feature importances. This paper advocates for the adoption of robust statistical methods as a superior alternative to model-derived feature importances, emphasizing three key advantages: the provision of directional information (ranging from -1 to +1), standardized comparison scales, and statistical validation through p-values. To enhance the reliability and interpretability of feature importance analysis, this paper introduces a comprehensive framework that incorporates five nonlinear, nonparametric statistical methods. This approach is designed to enhance the rigor and clarity of feature importance assessments in battery research and related fields.