Predicted Mean Vote (PMV) summarizes thermal response as a population mean, and many machine-learning studies likewise reduce thermal sensation votes (TSVs) to single-point regression outputs or unordered class labels. Yet operational decisions often depend on the full ordered distribution of TSV states, especially the warm and cold tails associated with discomfort risk. We therefore reframe TSV prediction as an ordinal learning problem and train cumulative-link models that produce class probabilities suitable for risk-aware decisions. Using the ASHRAE Global Thermal Comfort Database II and a large Chinese dataset, we harmonize features with a fold-safe, simple imputation policy (primarily Ta-Tr with an indicator) and evaluate 5-fold out-of-fold performance and cross-corpus swaps. Across metrics that capture accuracy and safety (MAE/Within-1, quadratic-weighted k, FAR@2, CRPS, ECE), the ordinal booster achieves the highest quadratic-weighted k(fewer long-distance errors). Overall FAR@2 is comparable across learning models on the aggregate, with nominal/neural baselines slightly lower and the ordinal model showing lower warm-tail but higher cold-tail FAR. Nominal baselines are best calibrated; post-hoc temperature scaling improves probability quality without changing rankings. Calibration and confusion diagnostics together with permutation-importance analyses indicate physically sensible effects. Cross- corpus tests show stable generalization across the two corpora. Ordinal learning with calibrated probabilities provides a transparent, reproducible path to risk-aware comfort prediction using existing field datasets.
Expand Image
