TOP: Taxi Destination Prediction Based on Trajectory Knowledge Graph

Destination prediction is an important issue for location based services (LBSs) and intelligent urban traffic planning. Existing destination methods either ignore spatial correlations or temporal features of the trajectories. We propose four spatial ontologies and four relations to construct the trajectory knowledge graph, such that the spatial explicit and implicit features are captured. Then, we propose a new destination prediction method TOP w.r.t the trajectory knowledge graph and the embedded trajectories. Specifically, the relational graph convolutional network is used on the trajectory knowledge graph to obtain the embedding representation of the spatial entities; trajectories are transformed into embedding trajectories utilizing the embedding representation of the spatial entities; GRU is employed to learn the temporal features on the embedded trajectories; finally, external metadata, such as workdays, holidays, and departure time are fused together to predict the destination. We also conduct a series of experiments on a real dataset. The experimental results proved the effectiveness of the proposed method.