Nonparametric Bayesian matrix factorization for assortative networks

We describe in detail the gamma process edge partition model that is well suited to analyze assortative relational networks. The model links the binary edges of an undirected and unweighted relational network with a latent factor model via the Bernoulli-Poisson link, and uses the gamma process to support a potentially infinite number of latent communities. The communities are allowed to overlap with each other, with a community's overlapping parts assumed to be more densely connected than its non-overlapping ones. The model is evaluated with synthetic data to illustrate its ability to model as-sortative networks and its restriction on modeling dissortative ones.