SARN: Spatial Structure-Aware Road Network Embedding via Graph Contrastive Learning

This is a pytorch implementation of the SARN paper:

@inproceedings{ChangTC023,
  author    = {Yanchuan Chang and Egemen Tanin and Xin Cao and Jianzhong Qi},
  title     = {Spatial Structure-Aware Road Network Embedding via Graph Contrastive Learning},
  booktitle = {Proceedings 26th International Conference on Extending Database Technology, {EDBT}},
  pages     = {144--156},
  year      = {2023},
}

Requirements

Ubuntu 20.04 LTS with Python 3.7.7
pip install -r requirements.txt
Download SF dataset here, and unzip -j SARN_dataset.zip -d data

Quick Start

First pre-train a SARN (cf. Section Self-supervised Training), then it can be used in downstream tasks, where the parameters can be fine-tuned or frozen (cf. Section Downstream Task Prediction).

Self-supervised Training

Pre-train a SARN model and evaluate its performance on road property prediction task with the frozen embeddings. The trained SARN and the corresponding learned embeddings are persisted to disk (in ./exp/snapshots/) for other downstream tasks. Logs are ouputted to the terminal and dumped to the log files in ./exp/log.

python train.py --task_encoder_model SARN --dataset SF

Downstream Task Prediction

We focus on three downstream tasks on road networks, including road property prediction, trajectory similarity prediction and shortest-path distance prediction - classify, trajsimi and spd for short, respectively.

Fine-tune the pre-trained SARN model and train a classify task model. (Prerequisites: a pre-trained SARN model, in other words, run the last command first).

python train.py --task_encoder_model SARN_ft --dataset SF --task_name classify --task_pretrained_model

Fine-tune the pre-trained SARN model and train a trajsimi task model.

python train.py --task_encoder_model SARN_ft --dataset SF --task_name trajsimi --task_pretrained_model

Fine-tune the pre-trained SARN model and train a spd task model.

python train.py --task_encoder_model SARN_ft --dataset SF --task_name spd --task_pretrained_model

Train a classify task model with the frozen embeddings.

python train.py --task_encoder_model SARN --dataset SF --task_name classify --task_pretrained_model

Train a trajsimi task model with the frozen embeddings.

python train.py --task_encoder_model SARN --dataset SF --task_name trajsimi --task_pretrained_model

Train a spd task model with the frozen embeddings.

python train.py --task_encoder_model SARN --dataset SF --task_name spd --task_pretrained_model

FAQ

Datasets

To use your own datasets, you may need to follow the steps below:

Download the OpenStreetMap xml dataset of a certain area. (See ./data/OSM_SanFrancisco_downtown_raw).
Extract the road network from the xml dataset and dump the data to the dedicated files. (See ./utils/osm2roadnetwork.py and ./data/OSM_SanFrancisco_downtown_raw).
Prepare the trajectory dataset used in trajsimi task. (See ./utils/traj_preprocess_sf.py).
The ground-truth labels used in downstream tasks are created during the downstream task training, but such process will be only executed once. After the first time, such labels will be read from files created during the first execution. (See ./task/classifier.py::Classifier.classifier_datasets, ./task/traj_simi_v2.py::TrajSimi.load_trajsimi_dataset and ./task/spd.py::SPD.get_spd_dict).

Contact

Email changyanchuan@gmail.com if you have any inquiry.