Graph neural network (GraphSAGE)

Graph neural networks (GNNs) work on data where items are connected, such as counties linked to their geographic neighbors. Each layer updates a node's representation by combining its own features with information passed from connected nodes.

GraphSAGE is a widely used GNN that samples and aggregates neighbor features, which lets it generalize to nodes it did not see during training. Two layers means each node draws on its neighbors and its neighbors' neighbors.

For geographic health data, this captures spatial spillover: a county's outcomes depend partly on the hospitals, roads and populations around it. Spatial cross-validation, for example holding out whole states, is needed to avoid overstating accuracy.