A personal comfort model is a new approach to thermal comfort modeling that predicts individuals' thermal comfort responses, instead of the average response of a large population. However, securing consistent occupant feedback for model development is challenging as the current methods of data collection rely on individuals' survey participation. We explored the use of a new type of feedback, occupants' heating and cooling behavior with a personal comfort system (PCS) for the development of personal comfort models to predict individuals' thermal preference. The model development draws from field data including PCS control behavior, environmental conditions and mechanical system settings collected from 38 occupants in an office building, and employs six machine learning algorithms. The results showed that (1) personal comfort models based on all field data produced the median accuracy of 0.73 among all subjects and improved predictive accuracy compared to conventional models (PMV, adaptive) which produced a median accuracy of 0.51; (2) the PMV and adaptive models produced individual comfort predictions only slightly better than random guessing under the relatively mild indoor environment observed in the field study; and (3) the models based on PCS control behavior produced the best prediction accuracy when individually assessing all categories of field data acquired in the study. We conclude that personal comfort models based on occupants' heating and cooling behavior can effectively predict individuals' thermal preference and can therefore be used in everyday comfort management to improve occupant satisfaction and energy use in buildings.