Fig. 3
From: Explaining classification performance and bias via network structure and sampling technique
RQ1: Estimation errors on small samples using random node sampling. These are the squared estimation errors (SE) of the conditional probabilities \(P_{maj|maj}\) (x-axis) and \(P_{min|min}\) (y-axis) learned from small samples (\(pseeds\le {{30}\%}\)) using random nodes on sparse networks (\(N={2000}, d=0.004\)) with different levels of homophily (rows) and class balance (columns). Mean ROCAUC scores within each type of network is shown as \(\overline{ROCAUC}\). In general we see that homophilic networks require lower estimation errors—especially within majorities—to achieve high performance (green)