Online Learning of User-Specific Destination Prediction Models

In this paper, we introduce and evaluate two different mechanisms for efficient online updating of user-specific destination prediction models. Although users can experience long periods of regular behavior during which it is possible to leverage the visitation time to learn a static user-specific model of transportation patterns, many users exhibit a substantial amount of variability in their travel patterns, either because their habits slowly change over time or they oscillate between several different routines. Our methods combat this problem by doing an online modification of the contribution of past data to account for this drift in user behavior. By learning model updates, our proposed mechanisms, Discount Factor updating and Dynamic Conditional Probability Table assignment, can improve on the prediction accuracy of the best non updating methods on two challenging location-based social networking datasets while remaining robust to the effects of missing check-in data.