Graph Neural Networks
A network that builds each node's representation by repeatedly combining it with messages gathered from its graph neighbors, learning from relationships, not just isolated rows.
Each layer gathers messages from a node's neighbors and combines them with the node's own state to produce its next representation, and stacking layers lets information travel further across the graph. Past roughly three to four layers, this repeated averaging tends to make every node's representation converge toward the same value, a failure called over-smoothing.
GCN, GraphSAGE, and GAT differ mainly in how they aggregate neighbor messages — fixed normalized averaging, sampled subsets for scale, or learned attention weights, respectively.