Comparing multiple networks, in a way that appropriately addresses nodal dependence, is a difficult task. It is complicated by the fact that many networks under comparison vary in size, the number of nodes in the network.
We consider Bayesian versions of latent space models for network data (Hoff et al. 2002), which model nodes who are closer together in a (typically Euclidean) latent social space as more likely to be tied. The resulting posterior predictive distributions can provide a (model-based) lens into how well a network of one size compares to a network of a different size. Building on past work, we examine how the geometry of the latent space adjusts this lens and can further illuminate network differences in these forecasted distributions. We demonstrate how these prediction scores can be used to infer interesting behavioral patterns in game play data from an online social science experiment that examines group innovation in a competitive atmosphere.
|