[Paper Review] Examining COVID-19 Forecasting using Spatio-Temporal Graph Neural Networks

Amol Kapoor, Xue Ben+5 more|arXiv (Cornell University)|Jul 6, 2020

Data-Driven Disease Surveillance34 references148 citations

TL;DR

This study uses a spatio-temporal graph neural network with mobility data to forecast US county COVID-19 daily cases, achieving a 6% RMSLE reduction and a small absolute improvement in Pearson correlation.

ABSTRACT

In this work, we examine a novel forecasting approach for COVID-19 case prediction that uses Graph Neural Networks and mobility data. In contrast to existing time series forecasting models, the proposed approach learns from a single large-scale spatio-temporal graph, where nodes represent the region-level human mobility, spatial edges represent the human mobility based inter-region connectivity, and temporal edges represent node features through time. We evaluate this approach on the US county level COVID-19 dataset, and demonstrate that the rich spatial and temporal information leveraged by the graph neural network allows the model to learn complex dynamics. We show a 6% reduction of RMSLE and an absolute Pearson Correlation improvement from 0.9978 to 0.998 compared to the best performing baseline models. This novel source of information combined with graph based deep learning approaches can be a powerful tool to understand the spread and evolution of COVID-19. We encourage others to further develop a novel modeling paradigm for infectious disease based on GNNs and high resolution mobility data.

Motivation & Objective

• Motivate improved forecasting of COVID-19 cases by leveraging inter-region mobility.
• Propose a unifying spatio-temporal graph framework that captures both spatial and temporal dependencies.
• Utilize real-time mobility datasets to inform inter-county connectivity and movement patterns.
• Evaluate forecasting performance against traditional time-series baselines.

Proposed method

• Construct a 100-county spatio-temporal graph with spatial edges weighted by mobility flows.
• Use a skip-connection-enhanced spectral graph convolutional network to aggregate information across up to two spatial hops and seven days of temporal context.
• Represent node features with an initial temporal embedding H0 derived from d-day history via an MLP.
• Predict future cases by applying an MLP to the final concatenated embedding Hs.
• Train with MSLE loss using Adam optimizer on data from Jan 1 onwards, with a specific train/test split.
• Compare against baselines including ARIMA, LSTM, Seq2Seq, and mobility-augmented variants.

Experimental results

Research questions

• RQ1 Can a spatio-temporal GNN leveraging inter-region mobility improve COVID-19 case forecasting over traditional time-series models?
• RQ2 How does incorporating mobility data influence predictive performance across US counties?
• RQ3 What is the relative performance of GNN-based forecasting versus standard baselines when mobility signals are included or excluded?

Key findings

• The GNN outperforms baselines on RMSLE and correlation, achieving best or second-best scores across metrics.
• Incorporating mobility data improves deep models, while ARIMA can degrade with changing mobility patterns.
• The top-performing GNN variant achieves an RMSLE of 0.0109 and Corr of 0.9980 on the main evaluation rows.
• Mobility-informed Seq2Seq and LSTM show gains over their no-mobility counterparts, but GNN consistently performs best.
• Overall, the GNN approach demonstrates the value of high-resolution mobility data in learning disease spread dynamics.