From taxi trajectories to transit networks: a two-stage DQN-GNN framework for large-scale urban routing

Purpose This study aims to propose a complete data-driven framework for the Transit Network Design Problem (TNDP) using large-scale taxi GPS trajectories to infer travel patterns and optimize transit routes. Design/methodology/approach The proposed framework consists of two stages. First, a convolutional weighted multi-objective stop selection (CWMSS) algorithm is introduced to identify high-demand stop areas by integrating passenger demand, coverage strength, and walking distance using a distance exponential decay mechanism. Second, a Deep Q-Network (DQN) combined with a Graph Neural Network (GNN) is developed. This model uses a fine-grained neighbor-extension strategy to incrementally construct transit routes, optimizing a tunable multi-objective cost function that balances passenger experience and operational cost. Findings Evaluations using a decade-long New York City (NYC) taxi dataset show the CWMSS algorithm improves Pickup and Drop-off Records (PDRs) and Effective Coverage Area (ECA) by 9.39% and 21.07% against state-of-the-art stop selection methods. Furthermore, the DQN-GNN model outperforms two advanced heuristics and two state-of-the-art Reinforcement Learning (RL) baselines in passenger-operator balanced and passenger-oriented scenarios. Friedman and Wilcoxon signed-rank tests indicate that the observed improvements are statistically significant (p < 0.05). Originality/value This research addresses the limitations of applying RL to TNDP on large-scale real-world networks. It contributes a novel stop selection algorithm using convolutional operations and a DQN-GNN model. This approach not only resolves the curse of dimensionality inherent in large-scale urban environments but also successfully designs transit networks that simultaneously improve passenger experience and lower operational costs.