Analysis of Drinking Water Feasibility Using Machine Learning and Explainable Artificial Intelligence
作者
Ryan Ferdinand Hajadi,Zulfany Erlisa Rasjid
标识
DOI:10.1109/incosst69426.2026.11709610
摘要
Access to safe drinking water is a global public-health priority, yet conventional laboratory-based assessment is slow, costly, and ill-suited to high-volume monitoring. Here we present a classification pipeline using machine learning with explainable artificial intelligence to assess the safety of drinking water. A publicly available dataset of 7,996 water samples having 20 physicochemical and biological attributes was employed to train and to analyze six classifiers, LightGBM, HistGradientBoosting, CatBoost, Random Forest, Extra Trees, and SVM-RBF. In order to avoid any data leakage, strict three-way split was enforced in training, validation, and test sets. HistGB achieved the highest test accuracy with 97.06% among individual models. The accuracy of 97.19% was achieved only through combining the top three models in a weighted soft-voting ensemble, but McNemar's exact test indicated this gain was not significant. Feature attribution through SHAP on LightGBM and LIME on selected predictions revealed aluminium, cadmium, and silver as the dominant predictors, a finding that aligns with WHO and EPA regulatory thresholds, lending external validity to the model's decision logic.