CMStatistics 2016: Start Registration
View Submission - CMStatistics
B0276
Title: Likelihood inference for large stochastic blockmodels with covariates Authors:  Yves Atchade - University of Michigan (United States)
George Michailidis - U of Florida (United States)
Sandipan Roy - University of Bath (United Kingdom) [presenting]
Abstract: A covariate blockmodeling framework is introduced which is in the class of blockmodels that has been widely used in analysing social networks. We introduce a model that captures observations coming from a stochastic blockmodel with certain number of covariates. We devise a novel algorithm based on case-control approximation of the log-likelihood along with a subsampling approach. Our algorithm is based on dividing the subsamples in several cores and then using a single communication among the cores after every iteration. In each core we use a Monte-Carlo EM type algorithm for parameter estimation and latent node label updates. We compare our method with some other methods available for community detection in blockmodels. We also provide an application of our algorithm to a real world network comprising a collection of Facebook profiles with few specific covariates.