Exploring Temporal Graph Neural Networks for Autoregressive Forecasting of COVID-19

This repo accompanies the paper "Exploring Temporal Graph Neural Networks for Autoregressive Forecasting of COVID-19" written by Skyler Wu '24 for Professor Melanie Weber's APMTH 220: Geometric Methods for Machine Learning, Spring 2024.

Raw data from the Google COVID-19 Open Data Repository as well shapefiles for map visualization from the US Census can be found in the raw folder, while processed data can be found in the processed folder (e.g., daily-aggregated-weekly data, central files for saving cleaned location_key and latitude/longitude data). I zipped the contents of the raw folder to comply with GitHub file size limits. The notebooks used to process the raw data and prototype some of the below pipelines can be found in notebooks.

Within the scripts folder:

• linear_main.py contains the main pipeline for running linear vector autoregression and standard autoregression models for COVID-19 disease forecasting.
• dcrnn_main.py contains the main pipeline for running DCRNN models for COVID-19 disease forecasting.
• tgcn_main.py contains the main pipeline for running T-GCN models for COVID-19 disease forecasting.

To run all experiments on the FASRC Cannon high-performance computing cluster (or another SLURM-based cluster): run bash {linear, dcrnn, tgcn}_main_runscript_driver.sh (select one of them) after making appropriate updates to the filepaths, fairshare accounts, and partition names, etc.

To process raw experimental results and generate figures: use analyzer.ipynb to generate aggregated log files and then use main_visualizer.ipynb to generate figures.