On Semiparametric Exponential Family Graphical Models

• Activity Number: 326
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Yang Ning* and Zhuoran Yang and Han Liu
• Companies: Princeton University and Tsinghua University and Princeton University
• Keywords: graphical model ; high dimensional inference ; nonconvex penalty ; semiparametric model

Abstract

We study a general class of semiparametric exponential family graphical models for the analysis of high dimensional mixed data. Unlike existing methods that require the nodewise conditional distributions to be fully specified by known generalized linear models, we allow the nodewise conditional distributions to be generalized linear models with unspecified base measures. For graph inference, we propose a new procedure named NOSI which is invariant to arbitrary base measures and attains optimal rates of convergence for parameter estimation. We also provide theoretical guarantees for both local and global inference of the true graph. Thorough numerical simulations and a real data example are provided to back up our results.