Learning Time-Based Presence Probabilities

Many potential pervasive computing applications could use predictions of when a person will be at a certain place. Using a survey and GPS data from 34 participants in 11 households, we develop and test algorithms for predicting when a person will be at home or away. We show that our participants’ self-reported home/away schedules are not very accurate, and we introduce a probabilistic home/away schedule computed from observed GPS data. The computation includes smoothing and a soft schedule template. We show how the probabilistic schedule outperforms both the self-reported schedule and an algorithm based on driving time. We also show how to combine our algorithm with the best part of the drive time algorithm for a slight boost in performance.