A Comparative Study of GNNs and Rule-Based Methods for Synthetic Social Network Generation

Abstract

Social network data, and web-graph data in general, sees great amounts of research and is widely used commercially. Due to privacy or other concerns it often cannot be shared verbatim. Synthetic counterparts to such data allow easier and safer sharing. Models currently applied to producing such synthetic data are typically rule-based, and though efficient, struggle to express complex topological characteristics. Graph Neural Network (GNN) models, in other domains such as molecules and proteins, trade-off efficiency for expressive abilities in generating graph structures. We compare GNNs and rule-based models for synthetic social networks and web-graphs from Facebook, Twitch, GitHub and Deezer. As a necessary data-processing step we propose and demonstrate a sampling procedure to construct network datasets from using datasets from single large networks. With structural and social network specific measurements, we evaluate how realistically synthetic networks behave in typical social network analysis applications. We find that the Gated Recurrent Attention Network (GRAN), an archetypal GNN model, extends well to social networks. In comparison to popular rule-based methods, R-MAT (Recursive-MATrix) and BTER (Block Two-level Erdos-Renyi), GRAN is better able to produce a variety of social networks and web-graphs.