Fig. 5
From: Explaining classification performance and bias via network structure and sampling technique
RQ3: Direction of bias. We measure the direction of bias by comparing the true positive rates of each class. The unbiased case is \(bias=0.5\), when both classes have the same true positive rates. Otherwise, results are biased towards majority nodes \(bias<0.5\), or towards minority nodes \(bias>0.5\). These are observations on small samples (\(pseeds\le {{30}\%}\)) where fluctuations are high, after this point bias scores converge or get better (i.e., towards \(bias=0.5\)). We see that classification on heterophilic networks is less biased than in neutral and homophilic networks