Fig. 2

The top row from left to right shows the (node) degree distribution, average (node) clustering coefficient per degree, and frequency of connected component sizes of five real-world networks. The bottom row shows the same quantities, but for three synthetic graph models: Erdös-Rényi (ER)(Erdös and Rényi 1959), Barabási-Albert (BA) (Albert and Barabási 2002), and Block Two-Level Erdös-Rényi (BTER) (Seshadhri et al. 2012). NOL-HTR is able to learn to increase network size on BTER and similar real-world networks. We find that degree, clustering coefficient, and size of the connected component are all relevant, as well as interpretable, features for learning