Semisupervised Graph Neural Networks for West Nile virus forecasting
Abstract
West Nile virus (WNV) is one of the most widespread vector-borne diseases in the United States today, and forecasting WNV prevalence has become an essential task for mosquito control programs. However, this remains a significant challenge due to WNV's complex transmission dynamics, involving an interacting system of environmental, meteorological, and geographic factors, and the limited quantity of data available. To meet this challenge, we propose a semi-supervised graph neural network (GNN) model that simultaneously learns from limited mosquito trap data and massive auxiliary environmental data. Our model is significantly larger than previous efforts, utilizes a more robust training regimen, and introduces a novel message-passing scheme that allows us to forecast even at out-of-graph locations. Experiments show that our model significantly outperforms existing fully supervised approaches in out-of-sample forecasting skill, particularly over longer time horizons (4-7 weeks).