Spatiotemporal Prediction of Chickenpox Cases using Graph Attention Network

Prediction of epidemiological systems is a classical spatial-temporal problem. As epidemiological processes evolve quickly and are spatially heterogeneous with high autocorrelation and seasonality, the problem is challenging on many grounds. Commonly used statistical and machine learning methods are limited in their effectiveness to understand the severity of the epidemiological process and delineate associated factors of spread. Due to great expressive power and convincing performance, Graph Neural Network approaches have attracted much interest in recent years. These approaches with Machine Learning methods focus on tasks such as node classification, link prediction, and clustering to reveal functional connectivity patterns of objects in the unique non-Euclidean data structure. This study proposed a model to handle the spatial-temporal forecasting problem using the graph attention network. The model’s performance on the Chickenpox Cases in Hungary dataset was compared against the forecasting performance of other well-established recurrent graph convolutional neural networks. It is seen from the results that the proposed model achieved the best forecasting performance.