Fig. 2
From: Explaining classification performance and bias via network structure and sampling technique
RQ1: Network structure, sample size and overall performance using random node sampling and network-only Bayes classifier (nBC). Classification performance, measured by ROCAUC scores, is shown on the y-axis for networks with \(N={2000}\) nodes and different levels of homophily (x-axis), edge density (rows), class balance (columns), and sample size (colors). Dots represent mean ROCAUC scores over 50 runs, and error bars their respective standard deviation. In general we see that: (i) neutral networks \(H=0.5\) cannot be classified better than a random classifier; (ii) heterophilic networks \(H<0.5\) require smaller samples to achieve high and stable classification performance compared to homophilic networks \(H>0.5\); (iii) Dense networks \(d=0.02\) achieve higher ROCAUC compared to sparse networks \(d=0.004\)