Approximate Conditional Inference for Degree-Corrected Network Models

• Activity Number: 254
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Daniel Klein*+
• Companies: Brown University
• Keywords: stochastic blockmodel ; conditional inference ; networks ; degree heterogeneity

Abstract

Stochastic blockmodels are commonly used to detect community structure in network data. Real-world networks often exhibit degree heterogeneity that makes the recovery of block structure difficult. Degree-corrected extensions have been proposed to model this heterogeneity, but at the cost of introducing an incidental parameter problem. In certain scaling regimes, e.g., bounded expected degree, efficiency and consistency for the estimation of a link covariate effect are impacted. We explore conditioning on the in- and out-degrees to eliminate the incidental parameters from the model. Efficient inference is made possible by using a sequential importance sampling algorithm to approximate the conditional likelihood surface. We characterize finite-sample performance and present some asymptotic results.