Spatial–Temporal Graph Learning for Taxi Origin–Destination Demand Prediction

ABSTRACT Accurate prediction of taxi origin–destination demand contributes to optimizing operational management, relieving traffic congestion, and improving the transportation capacity of urban road networks. To address the challenges of origin–destination semantic differentiation and data sparsity in taxi origin–destination demand prediction, we propose a multilevel continuous ‐ time dynamic node ‐ based attention network model (MCNAT). In the spatial–temporal graph construction unit, this model utilizes a dual‐strategy meta‐path random walk to achieve semantic differentiation and an attenuation function to update node representation. In the spatial–temporal correlation unit, this model employs multilevel spatiotemporal perception layers based on multi‐head attention mechanisms to aggregate local spaces and share parameter space. To mitigate the effects of data sparsity, L‐sparse is introduced in the prediction states to make the model more focused on edges with lower travel demand. We conducted simulations using two real datasets and compared MCNAT with other baselines. The results show that MCNAT outperforms the base model in all metrics, with MAE and RMSE improved by 11.11% and 11.87% compared to the optimal model, and the F 1 score metrics are above 0.69, which is improved by 11.1%.