Replicated Random Graphs under a Mixture of Low-Rank Decompositions with applications to Multiview Network Modeling

Abstract

Collections of network-valued data are prevalent in many scientific domains. Popular methods for analysis use averaging techniques to study statistical properties of the collection, leading to the possible loss of information or structure from the underlying distribution. In this paper we investigate a probabilistic framework for inferring global similarities and local deviations in a set of observed networks generated by a common random variable. This approach leverages a recently developed nonparametric Bayesian random graph model using a mixture of low-rank decompositions in order to facilitate both dimensionality reduction and clustering. We formulate the model and derive a Gibbs sampling procedure for posterior inference. Furthermore, we demonstrate a novel application for unsupervised learning in multivew networks.