AI-driven risk management for sustainable water distribution: A comparative study of resampling strategies and cost-sensitive predictive modeling for leakage failure

Aging water infrastructure and the resulting increase in pipe leaks pose significant operational and financial challenges for modern utilities, requiring more accurate tools for failure identification. This study presents a comprehensive benchmarking framework designed to predict pipe failure probability by evaluating a wide array of state-of-the-art classification models, including traditional baselines, tree-based ensembles, and emerging tabular deep learning architectures. The methodology integrates high-resolution datasets with a dedicated evaluation of spatially derived infrastructure indicators to capture the complex environmental and physical drivers of failure. To address inherent class imbalance, the study systematically benchmarks resampling strategies, such as SMOTE, ADASYN, and RUS, to determine whether these techniques truly improve decision-making performance. This assessment is grounded in the application of proper scoring rules, specifically Logarithmic Loss and Brier Score, alongside the introduction of the Area Under the Cost Curve to evaluate the economic implications of predictive performance across varying cost scenarios.