This is the official PyTorch implementation of the paper "Spatial–Temporal Synchronous Graph Transformer network (STSGT) for COVID-19 forecasting" that was presented at IEEE/ACM CHASE 2022 conference and published in Elsevier Smart Health Journal (2022).

• python >= 3.6
• pytorch >= 1.8.0

The following steps are required to replicate our work:

1. Download datasets.

• JHU Dataset - Download JHU COVID time-series data (download time_series_covid19_confirmed_US.csv for daily US infected cases and time_series_covid19_deaths_US.csv for daily US death cases) and save in data/COVID_JHU directory. This project used Mar 15,2020 - Nov 30,2021 for analysis.
• NYT Dataset - Download NYT COVID time-series data (download us-states.csv for daily US infected and death cases) and save in data/COVID_NYT directory. This project used Mar 18,2020 - Nov 30,2021 for analysis.

2. Generate Feature Matrix (X) and Adjacency Matrix (W) from downloaded datasets.

• JHU Dataset (US) - Inside the folder data/COVID_JHU, run the file Generate_51_states_X_W.py to generate X and W matrix for 50 states of US and Washington D.C. (51 nodes of graph).
• JHU Dataset (Michigan) - Inside the folder data/COVID_JHU, run the file Generate_51_states_X_W_Michigan.py to generate X and W matrix for 83 counties of the state of Michigan (83 nodes of graph).
• NYT Dataset (US) - Inside the folder data/COVID_NYT, run the file Generate_51_states_X_W_NYT.py to generate X matrix for 50 states of US and Washington D.C. (51 nodes of graph). We used the same adjacency matrix (W) as generated using JHU dataset.

3. Generate Train, Validation and Test datasets from the generated X matrix.

• We divided the entire dataset in chronological order with 80% training, 10% validation and 10% testing.
• Run the file generate_training_data.py to generate the processed files train.npz, val.npz, test.npz from X matrix and save the processed files in data/COVID_JHU/processed or data/COVID_NYT/processed. Use the confirmed or deaths in the argument to generate infected and death cases processed files respectively.

# For JHU Daily Infected cases data
python generate_training_data.py --traffic_df_filename "data/COVID_JHU/covid19_confirmed_US_51_states_X_matrix_final.csv" 

# For NYT Daily Death cases data
python generate_training_data.py --traffic_df_filename "data/COVID_NYT/covid19_NYT_deaths_US_51_states_X_matrix_final.csv"

1. Define paths and hyper-parameters in configuration files.

• Refer to the files config/COVID_JHU.conf and config/COVID_NYT.conf for the data paths, hyper-parameters and model configurations used for training and testing.
• The sensors_distance in the config files indicate the path to the adjacency matrix W.

2. Train the model

python train.py --epochs 100 --learning_rate 0.001 --expid 1 --print_every 20

1. The pre-trained models could be found in checkpoints/pretrained_models

• Refer to the required folder JHU or NYT, Infected or Deaths for infected or death cases respectively and our model is in folder STST

2. Test the model

• An example for testing with COVID_JHU dataset's daily infected cases and COVID_NYT dataset's daily death cases with our model STST (name in code for STSGT model) is given here. The ... _best_model.pth indicates the model with the lowest Mean Absolute Error (MAE) on the validation set.

# For JHU Daily Infected cases data with our trained model
python test.py --checkpoint "checkpoints/pretrained_models/JHU_States_Infected/STST/exp_2_1654.67_best_model.pth"

# For NYT Daily Death cases data with our trained model
python test.py --checkpoint "checkpoints/pretrained_models/NYT_States_Deaths/STST/exp_1_19.06_best_model.pth"

• Please choose the correct configuration file with the DATASET variable in both train.py and test.py.

Please cite our paper if you find this work useful for your research:

@article{banerjee2022spatial,
  title={Spatial--temporal synchronous graph transformer network (STSGT) for COVID-19 forecasting},
  author={Banerjee, Soumyanil and Dong, Ming and Shi, Weisong},
  journal={Smart Health},
  volume={26},
  pages={100348},
  year={2022},
  publisher={Elsevier}
}