Pitfalls of XAI interpretation in environmental modeling: A warning on model bias in air quality data analysis

Abstract

Jung et al. (2025) achieved high predictive accuracy in interpolating missing ozone data using graph machine learning (ML) and conducted feature importance analysis with explainable AI (XAI). This correspondence acknowledges their significant contribution but discusses the limitations and biases inherent in ML models and XAI methods (e.g., Random Forest/Bootstrap Test, SHapley Additive exPlanations (SHAP)) and their impact on the reliability of derived feature importance. High predictive accuracy does not necessarily guarantee trustworthy interpretation of feature relevance, as evidenced by inconsistent importance rankings across models and XAI techniques. To enhance interpretability and scientific reliability, we advocate a validation strategy integrating ML with rigorous statistical analysis. It combines model-driven insights with statistical measures such as Spearman's rho and Kendall's tau, and information-theoretic metrics like Mutual Information and Total Correlation to capture complex, non-linear dependencies. Such integration improves the robustness of feature importance assessments and supports more reliable interpretations in environmental modeling.