<![CDATA[This study proposes a robust and interpretable machine learning framework for water quality classification using a publicly available water quality dataset containing 7,996 samples and 20 physicochemical features with an imbalanced class distribution (88.59% majority and 11.41% minority). The study addresses the critical issue of biased classification toward the majority class, which can lead to risk-prone misclassification of unsafe water. An ensemble stacking model combining XGBoost, LightGBM, and CatBoost with a Random Forest meta-learner (passthrough) was developed using an anti-leakage pipeline integrating RobustScaler and SMOTE within stratified 80:20 train–test cross-validation, while hyperparameter tuning was optimized using F1-score to improve minority-class performance; SHAP was further applied for global and local explainability. The proposed model achieved an F1-score of 0.8563 for the minority class and a ROC-AUC of 0.9846, indicating strong discriminative performance, while SHAP analysis identified ammonia as the most influential feature and revealed that False Negative errors were mainly caused by complex feature interactions. The study contributes an integrated framework combining stacking ensemble learning, anti-leakage evaluation, and SHAP-based global–local interpretation to support more reliable and transparent water quality classification; however, the findings are currently limited to a single dataset and and require multi-dataset validation.]]>
