Evaluating water potability: a comparative analysis of machine and deep learning models for binary classification

Ensuring the safety of drinking water remains a global public health priority, which requires the development of robust data-driven predictive tools to augment traditional laboratory test-ing. This study presents a rigorous comparative evaluation of various Machine Learning (ML) and Deep Learning (DL) architectures to facilitate the binary classification of water potability. Using a comprehensive, publicly available open-source benchmark data set comprising 3,276 unique water quality profiles, the research methodology employed standardized preprocessing pipelines and stratified 5-fold cross-validation to ensure model generalizability and robustness. Our quantitative results demonstrate that non-linear models significantly outperform linear baselines, with the Support Vector Machine (SVM) utilizing a Radial Basis Function (RBF) kernel emerging as the top performer, achieving an accuracy of 0.6713 and an Area Under the Curve (AUC) of 0.6661. However, the analysis uncovers a critical “performance ceiling” inherent in the current feature set. All models exhibited a systemic bias characterized by high Preci-sion but notably low Recall (peaking at 0.3490), driven by significant class imbalance and weak feature-to-target separation within the raw chemical parameters. To resolve this, future deploy-ments must prioritize imbalance mitigation strategies, such as Synthetic Minority Over-sampling Technique (SMOTE), and advanced feature engineering. Incorporating these recommendations will allow water management authorities to transition from reactive testing to real-time preven-tative risk mitigation, ultimately safeguarding public health by reliably identifying hazardous water sources prior to distribution.