Fig. 6
From: Explaining classification performance and bias via network structure and sampling technique
Classification performance on empirical networks. We run the collective classification algorithm using random node sampling on five real-world networks: a) sexual contact network, b, c) university networks, d) reciprocal hypher-link network and e) developer-follower network. Properties of the networks are shown as H for homophily and B for class balance. ROCAUC scores using different sample sizes are shown on the y- and x-axis, respectively. Results from real networks are shown as “empirical” (dark blue), and their synthetic counterparts as “BA-Homophily” (light blue). Overall, results from synthetic networks follow a similar pattern as the empirical counterpart, except GitHub when samples are small