Critical evaluation of feature importance assessment in FFNN-based models for predicting Kamlet-Taft parameters

Abstract

Mohan et al. developed a feed-forward neural network (FFNN) model to predict Kamlet-Taft parameters using quantum chemically derived features, achieving notable predictive accuracy. However, this study raises concerns about conflating prediction accuracy with feature importance accuracy, as high R² and low root mean square error (RMSE) do not guarantee valid feature importance assessments. The reliance on SHapley Additive exPlanations (SHAP) for feature evaluation is problematic due to model-specific biases that could misrepresent true associations. A broader understanding of data distribution, statistical relationships, and significance testing through p-values is essential to rectify this. This paper advocates for employing robust statistical methods, like Spearman's correlation, to effectively assess genuine associations and mitigate biases in feature importance analysis.