ABSTRACT
In this paper, we analyze the performance of graph convolutional networks (GCNs) in predicting COVID-19 incidence in states and union territories (UTs) in India as a semi-supervised learning task. By training the model with data from a small number of states whose incidence is known, we analyze the accuracy in predicting incidence levels in the remaining states and UTs in India. We explore the effect of pre-existing factors such as foreign visitor count, senior citizen population and population density of states in predicting spread. To show the robustness of this model, we introduce a novel method to choose states for training that reduces bias through random sampling in five regions that cover India’s geography. We show that GCNs, on average, produce a 9% improvement in accuracy over the best performing non-graph-based model and discuss if the results are feasible for use in a real-world scenario. © 2022 IEEE.