Random Node Sampling
RNS reduces memory and runtime while often matching or improving full-graph test accuracy. RNS has one principal sampling hyperparameter, the number of parts m. Random Node Sampling partitions all nodes uniformly into approximately equal d…
1 sources - 5 claims
RNS reduces memory and runtime while often matching or improving full-graph test accuracy. RNS has one principal sampling hyperparameter, the number of parts m. Random Node Sampling partitions all nodes uniformly into approximately equal disjoint subsets at the start of each epoch. RNS uses the induced subgraph on each subset as a mini-batch and computes loss only for training nodes inside that batch. RNS is presented as a strong default baseline for scalable transductive GNN training.