Travel Destination Prediction Based on Origin-Destination Data

Most of the existing destination prediction methods are based on historical travel trajectory data. There is rarely method to predict users' travel destination only depending on departure time and the coordinates of departure point. In this paper, we use a real-world travel dataset, which only contains time and the coordinate of users' travel location, and no trajectories, we propose a new destination prediction algorithm, which is composed of three modules, including candidate destinations supplement, feature extraction and classifier training. For some users who have rarely travel records, according to a supplement rule, we choose tens of candidate destinations from millions of data. We extract statistical feature, temporal feature, spatial neighbor feature and graph feature from the perspective of the user group, time and geographical location. Finally, the performance of our proposed algorithm in terms of score and running time is demonstrated by experiments.