Leveraging graph neural networks and mobility data for COVID-19 forecasting

A team of Brazilian computer scientists has published new research demonstrating that Graph Neural Networks (GNNs) significantly outperform traditional forecasting models for predicting volatile COVID-19 case counts. The study, led by Fernando H. O. Duarte and four colleagues, analyzed mobility networks in Brazil and China to show that spatial relationships between locations—captured through GNN architectures like GCRN and GCLSTM—provide essential context that simple temporal models miss.

While standard Long Short-Term Memory (LSTM) networks performed adequately for smooth, cumulative case trends, they struggled with the volatility of daily predictions. The researchers found that GNNs achieved statistically significant improvements (Nemenyi test, p < 0.05) by incorporating how human movement between regions influences disease spread. A key innovation was structural sparsification of input graphs, where removing negligible connections between locations reduced predictive error and enhanced model stability.

The study addresses a longstanding debate in epidemiological modeling about whether complex spatio-temporal architectures provide meaningful advantages over simpler approaches. By framing the forecasting problem as both regression and binary classification tasks, the team demonstrated that context size and prediction horizon significantly impact model performance. Their work provides concrete evidence that mobility data, when properly structured through graph representations, offers critical predictive power for public health planning during pandemics.