STCF: Spatial-Temporal Contrasting for Fine-Grained Urban Flow Inference

Left: coarse-grained taxi flow map; Right: inferred fine-grained taxi flow map by STCF.

This repo provides a reference implementation of STCF framework described in the following paper:

Spatial-Temporal Contrasting for Fine-Grained Urban Flow Inference
Xovee Xu, Zhiyuan Wang, Qiang Gao, Ting Zhong, Bei Hui, Fan Zhou, and Goce Trajcevski
IEEE Transactions on Big Data, vol. 9, no. 6, pp. 1711--1125, Nov 2023. https://doi.org/10.1109/TBDATA.2023.3316471

Data

We use TaxiBJ P1-P4 and BikeNYC datasets.

You can download all five datasets at:

• Google Drive
• Baidu Drive (password: ndep)

Environment

[Update] Latest Tensorflow environment supported.
In this repo, STCF is implemented by Python 3.9, TensorFlow 2.9.1, cudatoolkit 11.2, and cudnn 8.1.0. Note: above environment requires NVIDIA driver version >450.80.02.

Create a virtual environment and install GPU-support packages via Anaconda:

# create virtual environment
conda create --name=stcf python=3.9
conda activate stcf
conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0

# install tensorflow==2.9.1, scikit-learn>=1.0
pip install -r requirements.txt

If tensorflow cannot identify cudatoolkit, try to configure the system path as shown in the 4. GPU setup section of this guide.

Usage

Step 1: Pre-train the spatial- and temporal-contrasting networks:

# spatial
python code/sc_pretrain.py --dataset taxi-bj/p1 --model sc 
# temporal
python code/tc_pretrain.py --dataset taxi-bj/p1 --model tc  

Step 2: Fine-tune the coupled network and evaluate performance:

# fine-tune
python code/stcf.py --dataset taxi-bj/p1 --sc sc --tc tc --model stcf
# evaluate
python code/evaluate.py --dataset taxi-bj/p1 --model stcf

More options are described in the code.

Cite

If you find our paper & code are helpful for your research, please consider citing us 💟

@article{xu2023stcf, 
  author = {Xovee Xu and Zhiyuan Wang and Qiang Gao and Ting Zhong and Bei Hui and Fan Zhou and Goce Trajcevski}, 
  title = {Spatial-Temporal Contrasting for Fine-Grained Urban Flow Inference}, 
  journal = {IEEE Transactions on Big Data}, 
  year = {2023},
  volume = {9},
  number = {6},
  pages = {1711--1725},
  doi = {10.1109/TBDATA.2023.3316471},
}

Acknowledgment

We are particularly grateful for the assistance given by Yuhao Liang and Ce Li. We would like to show our gratitude to the authors of UrbanFM, FODE, UrbanODE, and others, for sharing their data and codes. We express our gratitude to reviewers and editors for giving constructive feedbacks. This work was initially submitted to a conference in January, 2021.

Contact

For any questions, please open an issue or drop an email to: xovee.xu at gmail.com