Fig. 6
(Left) Radar plot of normalized feature importance scores for DO-input, T-input, turbidity-input, Cl-input, EC-input, pH-input, and NO₂-input obtained from the four methods (MDI, permutation importance, mutual information, and SHAP). (Right) Stacked bar chart showing how many times each input variable was assigned to a given rank (Rank 1 to Rank 7) across the four-feature selection and interpretability techniques. Higher bars in the upper ranks (e.g., Rank 1–2) indicate more consistent identification of a variable as important, whereas taller bars in lower ranks (e.g., Rank 6–7) reflect consistently low importance.
