Abstract
Replicated network data are available in many research fields. In connectomic applications, inter-connections among brain regions are collected from each patient, with interest focusing on studying common structures and individual differences in the architecture of interconnections in the brain. Current methods study shared connectivity patterns or reduce network data to summary statistics prior to statistical analysis, while we propose a fully generative Bayesian nonparametric approach for modeling the population distribution of network-valued data. The joint distribution of the edges follows a multivariate Bernoulli distribution, with transformed edge probability vectors expressed as the sum of a shared similarity vector and a class-specific deviation modeled via flexible mixtures of low-rank factorization exploiting the network structure. The formulation is provably flexible, leads to a simple Gibbs sampler, and provides a framework for clustering network-valued data, while adaptively shrinking towards lower dimensional representations. We discuss theoretical properties and illustrate the performance in simulations and application to human brain networks.

This publication has 32 references indexed in Scilit: