Personalized Prediction of Indoor Comfort Using Graph Convolutional Matrix Completion

Recent progress in environment sensing technology focuses more on measuring the physical properties of the environment, e.g., temperature and noise, but lacks the ability to understand subjective responses, or feelings about the environment, e.g., indoor comfort. Feelings depend on both environmental conditions and individual needs and preferences. Different people may feel differently in the same room experiencing the same conditions. In this work, we apply a crowdsensing based approach to predict personalized indoor comfort. We assume that similar users share similar feelings about comfort, and that indoor comfort is related to a fixed set of conditions, e.g., space, humidity, temperature. We surveyed existing users of a case study building and used their responses to learn how to predict the personal responses of new users. Technically, we apply a graph convolutional matrix completion (GC-MC) method to predict the comfort of other users, by learning the dependency between the user profiles and their ratings to a fixed set of survey questions. We collect a kitchen survey dataset of 59 questions and in total 29 users of diverse profiles.