Personalised Comfort Metrology: Measuring the Reliability of Augmented Thermal Sensation Vote Data

Modern smart buildings use Personalized Comfort Models (PCMs) to balance energy efficiency and occupant comfort based on user feedback. However, their effectiveness is limited because most feedback data is missing 75-96% due to occupants getting tired of frequent surveys. This study evaluates 16 univariate interpolation algorithms-including two ordinal-native comparators-to reliably impute missing categorical Thermal Sensation Votes (TSV) without distorting statistical distributions or environmental correlations. Using data from 10 apartments, all continuous imputations were subjected to discrete bounding and post-interpolation rounding to match valid ASHRAE integers. Performance was assessed through Pearson and Spearman correlation structure preservation and three distribution tests. Results show that discrete rounding effectively neutralises non-physical artefacts. The linear family and the ordinal-native Ordinal Step method achieved identical lowest mean absolute correlation perturbation (0.113), while zero-order hold proved the most destructive (1.81 × more accumulated error). Spearman and Pearson rankings were significantly consistent (Kendall's τ=0.69, p<0.001), validating the use of Pearson as a relative metric for ordinal TSV. Ultimately, bounded and rounded linear or cubic interpolations offer a computationally lightweight and robust method to reconstruct highly sparse subjective feedback, ensuring continuous PCM training despite fluctuating occupant participation.