Real time location prediction with taxi-GPS data streams

The prediction of the destination location at the time of pickup is an important problem with potential for substantial impact on the efficiency of a GPS-enabled taxi service. While this problem has been explored earlier in the batch data set-up, we propose in this paper new solutions in the streaming data set-up. We examine four incremental learning methods using a damped window model namely, Multivariate multiple regression, Spherical-spherical regression, Randomized spherical K-NN regression and an Ensemble of these methods for their effectiveness in solving the destination prediction problem. The performance of these methods on several large datasets are evaluated using suitably chosen metrics and they were also compared with some other existing methods. We found that the Multivariate multiple regression method has the best performance in terms of prediction accuracy but the Spherical-spherical regression method is the best performer when we take into account the accuracy time trade-off criterion. The next pickup location problem, where we are interested in predicting the next pickup location for a taxi given the dropoff location coordinates of the previous trip as input is also considered and the aforementioned methods are examined for their suitability using real world datasets. As in the case of destination prediction problem, here also we find that the Multivariate multiple regression method gives better performance than the rest when we consider prediction accuracy but the Spherical-spherical regression method is the best performer when the accuracy-time trade-off criterion is taken into account.