Contextualized Spatial–Temporal Graph Networks for Taxi Demand Prediction

Taxi demand prediction is a basis task for taxi on-demand services. However, accurate taxi demand prediction is still a challenge, because it needs to decouple complex time and space dependence as much as possible. Recently, the research on spatial-temporal dependence capturing in transportation field tends to adopt the method of combining graph neural network and time series models. However, these models expose some shortcomings, including the overlook of long-range temporal dependencies and dynamic spatial dependence. This paper presents an innovative graph model that is integrated with contextualized spatial-temporal information to improve taxi demand prediction. The spatial context refers to the spatial relationship between nodes and surrounding nodes. Temporal context refers to the temporal relationship between events. The experiments on a large-scale real taxi demand dataset demonstrates that the proposed model achieves the best results compared with state-of-the-arts methods.