A Graph Based Deep Learning Framework for Predicting Spatio-Temporal Vaccine Hesitancy

Abstract

Predicting vaccine hesitancy at a fine spatial level assists local policymakers in taking timely action. Vaccine hesitancy is a heterogeneous phenomenon that has a spatial and temporal aspect. This paper proposes a deep learning framework that combines graph neural networks (GNNs) with sequence module to forecast vaccine hesitancy at a higher spatial resolution. This integrated framework only uses population demographic data with historical vaccine hesitancy data. The GNN learns the spatial cross-regional demographic signals, and the sequence module catches the temporal dynamics by leveraging historical data. We formulate the problem on a weighted graph, where nodes are zip codes and edges are generated using three distinct mechanisms: 1) adjacent graph - if two zip codes have a shared boundary, they will form an edge between them; 2) distance-based graph - every pair of zip codes are connected with an edge having a weight that is a function of centroid distances, and 3) mobility graph - edges represent the number of contacts between any two zip codes, where the contacts are derived from an activity-based social contact network. Our framework effectively predicts the spatio-temporal dynamics of vaccine hesitancy at the zip-code level when the mobility network is used to formulate the graph. Experiments on the real-world vaccine hesitancy data from the All-Payer Claims Database (APCD) show that our framework can outperform a range of baselines.