Predicting regulatory violations in public drinking water systems: A data-driven approach

Enforcement of the Safe Drinking Water Act helps in protecting public health, yet agencies tasked with enforcement are besieged with numerous violations. In this research, a supervised machine learning approach was employed to predict the final regulatory action ('Resolved' or 'Archived') of 7,268 past drinking water violations within the U.S. Environmental Protection Agency's SDWIS database. Logistic Regression, Random Forest, and Gradient Boosting models were tuned using hyperparameter tuning and then trained on the data. The human-tuned Gradient Boosting model performed the best in prediction, with 82.3% accuracy and an F1-Score of 0.87 on the test set. Feature importance analysis revealed that the violation of the specific regulatory rule was the key predictor, followed by the quantitative extent of the violation (exceedance ratio) and the contaminant type. Findings reveal that the fate of drinking water violations is highly predictable based on their original characteristics. The research offers a proof-of-concept for predictive analytics as an environmental governance tool through which regulatory agencies may filter and prioritize high-priority violations to be addressed on a targeted basis.