We introduce a new family of network models, called hierarchical network models, that allow to represent in an explicit manner the stochastic dependence among the edges. In particular, each member of this family can be associated with a graphical model defining conditional independence clauses among the edges of the network, called the dependency graph. Every network model of dyadic independence assumption can be generalized to construct members of this new family. Using this new framework, we generalize the Erdös-Rényi and β-models to create hierarchical Erdös-Rényi and β-models. We describe various methods for parameter estimation as well as simulation studies for models with sparse dependency graphs. We also provide a comprehensive discussion on open problems related to these newly defined models. 1 arXiv:1605.04565v1 [stat.ME] 15 May 2016There is a natural duality between networks and dependency graphs: edges of the network are (binary) random variables, and hence can be considered to be nodes of a dependency graph. Edges of the dependency graph would then determine the conditional independence among these variables, i.e. the independence structure among the edges of the network.The first, and possibly the only, place where this duality has been used is in Frank and Strauss (1986), where they assumed the dependency graph to be the line graph of the complete graph, as defined in the graph theory literature. However, there have recently been other approaches to deal with certain "local" types of dependency in networks by using nodal attributes, i.e. extra information on the individuals presented by nodes of the network; see, for example, Schweinberger and Handcock (2012) and Fellows and Handcock (2012).The main goal of this paper is to exploit this duality between networks and dependency graphs. We basically use the model associated with a given dependency graph, and combine it with known network models that assume dyadic independence by further parametrizing the graphical model using the parameters in the network model. This allows us to derive new network models that preserve the independence structure of the given dependency graph and has similar properties to those of the original network model. For this purpose, we work with undirected graphical models, which imply that our proposed network models are of linear exponential family form.
3In this paper, we demonstrate our approach for three well-known classes of network models: The Erdös-Rényi models, defined by Erdös and Rényi (1959) and studied vastly in the literature of networks and random graph theory, the β-models, defined by Blitzstein and Diaconis (2010); Chatterjee et al. (2011) and recently studied by Rinaldo et al. (2013) for undirected networks, and the p 1 models defined by Holland and Leinhardt (1981) for directed networks.We also provide a method based on the gradient decent algorithm to estimate the maximum of the likelihood function for hierarchical Erdös-Rényi models. In principle, this method can be generalized to other families of hierarc...