Abstract In this paper we present a methodology to map individual occupants' thermal preference votes and indoor environmental variables into personalized preference models. Our modeling approach includes a new Bayesian classification and inference algorithm that incorporates hidden parameters and informative priors to account for the uncertainty associated with variables that are noisy or difficult to measure (unobserved) in real buildings (for example, the metabolic rate, air speed and occupants’ clothing level). To demonstrate our approach, we conducted an experimental study in private offices by considering thermal comfort delivery conditions that are representative of typical office buildings. Personalized preference models were developed with the training dataset and the developed algorithms were used in a detailed validation process. The proposed model showed better prediction performance compared to previous methods. Towards realization of preference-based control systems, this study also addresses practical limitations associated with controlling model complexity and data efficiency as well as using effective model evaluation metrics to train reliable personalized preference models in the real world.