A Popularity Scaled Latent Space Model for Large-Scale Directed Social Network

Abstract

Large-scale directed social network data often involve degree heterogeneity, reciprocity, and transitivity properties. A sensible network generating model should take these features into consideration. To this end, we propose a popularity scaled latent space model for the large-scale directed network structure formulation. It assumes for each node a position in a hypothetically assumed latent space. Then, the nodes close (far away) to each other should have larger (less) probability to be connected. As a consequence, the reciprocity and transitivity properties can be analytically derived. In addition to that, we assume for each node a popularity parameter. Those nodes with larger (smaller) popularity are more (less) likely to be followed by other nodes. By assuming different distributions for popularity parameters, different types of degree heterogeneity can be modeled. Furthermore, based on the proposed model, a comprehensive probabilistic index is constructed for link prediction. Its finite sample performance is demonstrated by extensive simulation studies and a Sina Weibo (a Twitter-type social network in China) dataset. The performances are competitive.