DiffRoute: An End-to-End Framework for Multi-Step Delivery Route Prediction

Last-mile delivery route prediction is a critical task in intelligent logistics; however, traditional optimization methods and standalone deep learning models encounter significant challenges in terms of scalability, adaptability, and long-term sequence prediction. To address these issues, we propose DiffRoute, an integrated framework that combines Graph Attention Networks (GAT) for spatial modeling and iTransformer for temporal sequence processing. Our framework introduces delivery-specific adaptations, including Area of Interest (AOI) feature integration, dynamic route masking, and Gumbel-Softmax based differentiable path selection, enabling end-to-end optimization under discrete routing constraints. In addition, a frequent pattern attention module retrieves and aggregates historical trajectories to enhance performance in data-sparse scenarios. Experiments on a large-scale LaDe dataset demonstrate that DiffRoute consistently outperforms traditional heuristics, machine learning baselines, and deep learning models, achieving superior accuracy across the Hit Rate at 3 (HR@3), Kendall Rank Correlation (KRC), Location Shift Distance (LSD), and Euclidean Distance (ED) metrics. On the LaDe Shanghai dataset, DiffRoute achieves HR@3 = 74.32% and KRC = 76.23. On the LaDe Chongqing dataset reaching HR@3 = 75.21% and KRC = 77.56, In both Shanghai and Chongqing, DiffRoute achieves lower LSD and ED values than all baseline methods. Ablation studies further confirm the indispensable contributions of each module, whereas visualization analyses highlight the model’s interpretability and robustness. These results validated the effectiveness of DiffRoute as a model for route planning in current urban logistics.