Improving Grid-Based Location Prediction Algorithms by Speed and Direction Based Boosting

Grid-based location prediction algorithms are widely researched and evaluated. These algorithms usually integrate the speed and the direction during the learning process as regular contextual features, for example, like the time of the day or the day of the week. Unfortunately, the way speed and direction are currently used does not fulfill their potential. In this paper, we propose an alternative approach for integrating the user's current speed and direction in a post-processing mechanism that highly improves the algorithms' accuracy. We dynamically update the probabilities of the predictions provided by the existing (base) algorithms by dividing the surface into four areas while boosting the probabilities in some areas and reducing the probabilities in others. We evaluated our method on three well-known grid-based location prediction algorithms and two different datasets and were able to show that our method improves the predictions made solely by the algorithms. Our improvement was stable during the entire experiment for long-term predictions and for greater prediction distances, particularly in the cold start phase which is considered more difficult to improve.