Urban flow generation

The official PyTorch implementation of "Towards Generative Modeling of Urban Flow through Knowledge-enhanced Denoising Diffusion" (SIGSPATIAL'23). (PDF on arxiv)

NYC, Washington D.C. (DC), and Baltimore (BM) datasets are included.

The code is tested under a Linux desktop with torch 1.9.0 and Python 3.8.10.

Installation

Environment

Tested OS: Linux
Python >= 3.8
PyTorch == 1.9.0
torch_geometric == 1.7.2

Dependencies

Install PyTorch 1.9.0 with the correct CUDA version.
Use the pip install -r requirements.txt command to install all of the Python modules and packages used in this project.

Usage

Step-1 Pretrain to get KG embeddings:

bash pretrain.sh

Step-2 Train diffusion model:

bash train.sh

Step-3 Generate urban flow:

bash sample.sh

Step-4 Evaluate generated flow:

python evaluate.py

(The default dataset is NYC, you can modify the dataset in pretrain.sh, train.sh, sample.sh, and evaluate.py)

More Related Works

Hierarchical Knowledge Graph Learning Enabled Socioeconomic Indicator Prediction in Location-Based Social Network (WWW '23)

Note

The implemention is based on DDPM.

If you found this library useful in your research, please consider citing:

@inproceedings{zhou2023towards,
  title={Towards Generative Modeling of Urban Flow through Knowledge-enhanced Denoising Diffusion},
  author={Zhou, Zhilun and Ding, Jingtao and Liu, Yu and Jin, Depeng and Li, Yong},
  booktitle={Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems},
  pages={1--12},
  year={2023}
}