Mini-batch GNN Training

Graph mini-batching should be understood as changing the training objective rather than merely approximating full-graph training. Mini-batch training for GNNs changes the topology available to the model by removing edges that cross the batch boundary. Sampler choice affects the effective objective through both sampled-loss bias and mini-batch gradient variance. Subgraph mini-batch training computes the supervised loss only on training nodes inside the sampled subgraph.