Assessing supervised machine learning practice in urban water networks: A critical review of methodological transparency, reproducibility and reporting

Le résumé fourni par la source

In recent years, the advantages of machine learning (ML) have been clearly demonstrated in research on urban water infrastructure (UWI) and has been applied in a wide range of applications. This review critically assesses the current quality of ML implementations in UWI by examining >100 articles from the recent literature, with a particular focus on common pitfalls throughout the development process. Most reviewed articles placed strong emphasis on performance benchmarking but provided limited reporting on key ML methodology implementation and deployment. Only around one third of the reviewed articles documented essential tasks such as feature scaling or automatic hyperparameter optimisation, despite their importance for performance and generalisation. Additionally, fewer than 25% reported explainability and uncertainty quantification techniques, making reported performance gains difficult to explain or operationalise. Furthermore, the lack of standardised documentation makes the extraction of relevant information about the methods, workflows, and steps needed for reproducibility difficult. Together, these issues negatively affect the reproducibility and reduce comparison and trust in the developed approaches. This is particularly important, as trust and confidence in ML-based decisions are key requirements for successful transformation of research into practice. To address these issues, a methodological and reporting checklist is provided to guide the design and integration of ML applications in UWI throughout the development process and to highlight common pitfalls. The review and the checklist can support developers, technicians, and operators in future ML applications, helping to raise awareness on reproducible ML implementations in UWI applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.