Mini-batch graph neural network (GNN) training samples an 𝐿-hop neighborhood around supervised seed nodes at each step. When seeds are chosen uniformly their neighborhoods are largely disjoint, producing redundant message passing and inflating wall-clock time. We propose CASS (Community Aware Seed Selection), a drop-in modification that changes only how seeds are selected, leaving the neighborhood sampler, fanout schedule, model, and loss unchanged. CASS partitions the graph into connectivity-based communities (bounded-BFS weakly connected components) and samples seeds from one or afewcommunities perstep. Onthreereal-worldgraphs(YelpChi, Reddit, Elliptic), CASSreducestraining time by 14.6%–45.6% at comparable ranking and classification quality. We further show the speedup correlates with a simple, cheaply computed structural statistic—the 1-hop Jaccard neighbor overlap— giving a pretraining diagnostic for when CASS will help.