GNN Vaccination Model Cuts Peak Infections on Email-Eu-core Network in Preprint Simulations

The study models epidemic spread on the Email-Eu-core network, which encodes real email interactions as a multi-layer contact graph. Researchers trained graph neural networks and reinforcement learning agents to select vaccination targets, then compared outcomes against classical centrality metrics in 30 stochastic SIR simulations. This setup directly incorporates population heterogeneity and higher-order network motifs that mass-vaccination models ignore.

Results indicate the GNN policy identified structurally critical nodes missed by degree or betweenness centrality, lowering peak prevalence and epidemic duration. The dense connectivity and modest community structure of the tested network explain why traditional heuristics performed similarly to one another. These patterns suggest learning-based targeting can improve equity by protecting bridge populations that standard metrics undervalue, with direct implications for future pandemic preparedness planning.

Key limitations include reliance on a single network and purely simulation-based evaluation without empirical validation. Larger, multi-city contact datasets and prospective trials would strengthen evidence. The work aligns with emerging literature on network epidemiology that emphasizes relational structure over aggregate risk groups.