Fig. 5

Effect of node attribute-based noise on real datasets. Similar to Fig. 3 we plot the influence of the retain parameter \(\rho\) on networks with different strength of homophily. Networks are sorted by modularity of their label partition which can be seen as a proxy for homophily, if the connectivity of the network is indeed driven by homophily. However, the networks here do not have a one-to-one correspondence with the synthetic networks regarding minority size and homophily. Nonetheless we can see that similar effects are present in real and synthetic world networks, with dblp being the notable exception where some saturation happens already at around \(\rho =0.7\)