Abstract
Vehicle dispatching is essential for maintaining supply–demand balance in ride-hailing systems, but existing frameworks separate demand prediction from vehicle allocation, causing a mismatch between prediction accuracy and operational performance. This study proposes a decision-focused end-to-end learning framework that integrates spatiotemporal demand representation and vehicle-to-region dispatching within a unified trainable architecture. The model combines a multi-graph spatiotemporal prediction module with a recurrent Transformer-based dispatching module that iteratively refines vehicle assignments under capacity-aware constraints. A structured loss jointly considers demand prediction, dispatch objectives, and constraint violations, reducing dependence on expensive optimization labels. Experiments on real ride-hailing datasets from Chengdu, Xi’an, and Haikou show competitive dispatching quality and zero capacity-constraint violation rate (CCVR) in the reported city-level comparison. Relative to the Kuhn–Munkres benchmark, the proposed model reduces average online computation time from 242.47 s to 2.47 s, a 98.98% reduction (approximately 98.2-fold acceleration). The decision-oriented mechanism yields operationally useful demand representations, suggesting neural dispatchers offer a scalable approximation for high-frequency operations where responsiveness and feasibility are critical.