Effect of disregarding small clusters. MNIST dataset as a function of the cluster size K below which clusters are disregarded. HLC (k=15, edges with highest betweenness centrality have been discarded): solid lines (unweighted, red; weighted, blue). Phenograph (k=40): dashed lines (unweighted, red; weighted, blue). Inset: HLC clustering (k=15) after betweenness centrality clipping, with clusters smaller than K=80 displayed white.