Structure-aware Samplers

RNS achieves the best test accuracy among the evaluated mini-batch methods on OGBN-ARXIV, OGBN-PRODUCTS, and POKEC. The sampler comparison evaluates Neighborhood Sampling, ClusterGCN, GraphSAINT, LADIES, RNS, and full-graph training on three main datasets. The RNS node-subsampling bias argument does not apply to ClusterGCN or GraphSAINT because their sampled target sets are not conditionally uniform over training nodes. Structure-aware samplers may create batches that differ systematically from one another and lead to larger per-batch gradient variance. Prior scalable GNN methods often use structure-aware samplers to preserve local connectivity or reduce embedding variance.