cassia151 / raildelays-public

reference for thesis

Railway Delay Prediction with Spatial-Temporal Graph Convolutional Networks

Related Materials

Read the paper

Youtube (Presented at ITSC 2020)

UIUC Aerospace Engineering Department Coverage

Cite Us

If you found these materials helpful, please cite us in your work

@inproceedings{heglund2020railway,
  title={Railway Delay Prediction with Spatial-Temporal Graph Convolutional Networks},
  author={Heglund, Jacob SW and Taleongpong, Panukorn and Hu, Simon and Tran, Huy T},
  booktitle={2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC)},
  pages={1--6},
  year={2020},
  organization={IEEE}
}

Reproducing the Results

Setup

Note: The computer we use to run this model runs Ubuntu 18.04 and uses CUDA V9.1 to facilitate model training using a GPU. While this code may work with other configurations, we did not test other configurations during our development.

1. Clone this repository to your local machine

2. Create a data folder in the project base directory with the following subdirectories

   • ./data/raw
   • ./data/interim
   • ./data/processed

3. Download and unzip the raw data from this Google Drive link to ./data/raw

4. Create and activate the raildelays conda environment using the following command

    $ conda env create -f environment.yml
   

5. Run jupyter notebook in the notebooks directory

    $ cd ./notebooks
    $ jupyter notebook
   

   • Open the jupyter notebook in a browser, and open the "create_data" notebook
   • Change the base_dir in the first cell to the base directory for this repository
   • Run all the cells in "Imports" and "Link-Based Node Formulation" sections. This creates the input dataset for the STGCN model.

Running Models

1. Open the src/main.py and src/model_comparison.py and change the base dir to the raildelays base directory

2. Run the STGCN model using the following command

    $ cd src
    $ python main.py --n_timesteps_in=6 --n_timesteps_future=1
   

where n_timesteps_in can take values {6, 12} and n_timesteps_future can take values {1, 3, 6} for the experiments presented in the paper. This trains the model for a default of 25 epochs. The accuracy metrics are shown in the terminal after inference.

1. Run the comparison models using the following command

    $ python model_comparison.py --n_timesteps_in=6 --n_timesteps_future=1 --model_type=MLP
   

where n_timesteps_in can take values {6, 12}, n_timesteps_future can take values {1, 3, 6}, model_type can be {LR, MLP} for the experiments presented in the paper. The accuracy metrics are shown in the terminal after inference.