Upgrading UTCI through supervised learning using the RUROS dataset

Outdoor thermal comfort (OTC) plays a key role in climate-resilient urban planning, but widely used indices such as the Universal Thermal Climate Index (UTCI) are constrained by their complex thermophysiological formulations and limited interpretability. This paper proposes an explainable Machine Learning (ML) framework for binary OTC classification based on the RUROS dataset, which contains 6,079 valid questionnaire responses from seven European cities. By substituting thermoregulatory simulations with a data-driven methodology, this study offers a transparent tool that clarifies how environmental variables shape human thermal perception. We assessed 36 distinct workflows in seven cities, comparing inherently interpretable models such as Logistic Regression and Decision Trees against standard baselines. By validating on cities that were held out during training, we confirmed that the model generalizes well to varied climatic conditions. Our final recommendation, a Naïve Bayes model with Downsampling (NBD), relies on air temperature, wind speed, and relative humidity to estimate comfort probabilities. The NBD model outperforms the UTCI, reaching a balanced accuracy of 0.611 versus 0.585 for the index. Statistical verification using Wilcoxon signed-rank tests and bootstrap confidence intervals shows that this advantage is both consistent and meaningful. In addition to higher predictive accuracy, NBD provides a probabilistic framework that enables urban planners to map heat-risk areas in terms of likelihood rather than relying on fixed thresholds. Together, these results underscore the promise of explainable ML for delivering more flexible and dependable evaluations to support well-being in public urban environments.