We propose a Multi-Relational Graph-based Route Prediction (MRGRP) method, which enables fine-grained modeling of the correlations among tasks influencing couriers' decision-making and achieves accurate prediction. We encode spatial and temporal proximity, along with the pickup-delivery relationships of tasks, into a multi-relational graph, then design a GraphFormer architecture to capture these complex correlations. Furthermore, we introduce a route decoder that leverages information about couriers as well as dynamic distance and time contexts for personalized prediction. It also utilizes existing route solutions as a reference to find better outcomes.
- Tested OS: Linux
- Python >= 3.10.0
- PyTorch >= 1.7.0
- hydra == 1.1
python run.py --dataset_name <your_data_folder>
The metrics include KRC, LSD, ED, SR@k, HR@k, [email protected], LSD, and ED quantify the similarity between the predicted route and the ground truth from the global perspective, while SR@k, HR@k, and ACC@k measure their local similarity. Higher KRC, HR@k, ACC@k, SR@k, and lower LSD, ED indicate better performance of methods.
We illustrate the results of hyperparameter studies as follows.
From the results, we find that the hyperparameter settings are:
- # MRGC Layers: 3
- Task embedding sizes: 256
- Relative weight: 0.1
The diagram of our proposed multi-relational graph is as follows.
We illustrate the process of route reference encoding generation, which is employed in the decoder, as follows.
The original data is coming soon after undergoing the necessary de-identification and anonymization processes to ensure compliance with privacy regulations and ethical standards.