GE-GAN: Graph Embedding GAN for Road Traffic State Estimation

Official open-source implementation of:

Dongwei Xu, Chenchen Wei, Peng Peng, Qi Xuan, and Haifeng Guo.

GE-GAN: A novel deep learning framework for road traffic state estimation.

Transportation Research Part C: Emerging Technologies, 117, 102635, 2020.

Read the paper

Overview

GE-GAN estimates missing road traffic states by combining graph embedding and a generative adversarial network:

1. DeepWalk learns representations of the detector road network.
2. The learned representation selects detectors that are most relevant to a target detector.
3. A Wasserstein GAN learns the traffic-state distribution and estimates the target detector's state.

The paper evaluates the framework on Caltrans PeMS District 7 traffic volume data and a Seattle traffic-speed dataset.

Repository

GE_GAN/
├── data/                 Published PeMS and Seattle matrices and graph files
├── deepwalk.py           Graph embedding
├── get_data.py           Related-detector selection and data preparation
├── model.py              GE-GAN model
├── main.py               Experiment entry point and parameters
└── visualization.py      Result visualization

The implementation is the original research release and uses its historical TensorFlow-era dependency stack. For reproducible work, isolate the environment and record exact dependency versions.

Download the Caltrans PeMS source data

The standalone Caltrans PeMS Data Downloader can prepare the District 7 source period and export matrices compatible with this repository:

pip install git+https://github.com/wcc961129/pems-data-downloader.git
playwright install chromium

pems-data auth
pems-data fetch \
  --profile ge-gan-d7-2014 \
  --output data/ge-gan-d7-2014

The profile covers May 1 through June 30, 2014 and uses the 23-station header published in this repository. It creates:

ge_gan/combine_E_workday_n0.csv
ge_gan/combine_E_weekend_n0.csv

This workflow prepares the same PeMS source period and compatible matrix orientation. It does not claim byte-identical historical files or a complete end-to-end reproduction of every paper result. See the downloader's GE-GAN reproducibility notes.

Run the original experiment

From the repository root, review the paths and parameters in GE_GAN/main.py, then run:

python -m GE_GAN.main

Important configuration includes:

• data_flag: PeMS or Seattle.
• file_name: workday or weekend matrix.
• target_segement: target detector column.
• sliding_windows, epoch, batch_size: training parameters.
• select_nums, walk_length, window_size: graph embedding parameters.

Citation

If this repository contributes to your research, please cite:

@article{xu2020gegan,
  title={GE-GAN: A novel deep learning framework for road traffic state estimation},
  author={Xu, Dongwei and Wei, Chenchen and Peng, Peng and Xuan, Qi and Guo, Haifeng},
  journal={Transportation Research Part C: Emerging Technologies},
  volume={117},
  pages={102635},
  year={2020},
  doi={10.1016/j.trc.2020.102635}
}

If you use the separate PeMS downloader, cite its software release as well.

Links

• Caltrans PeMS Data Downloader
• Chenchen Wei's CSDN blog
• Paper on ScienceDirect

License

This project is released under the Apache License 2.0.