|
Activity Number:
|
432
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
IMS
|
| Abstract - #305402 |
|
Title:
|
Robust Inference for Sparse Graphical Models Using a Multivariate t Distribution and Penalized Likelihoods
|
|
Author(s):
|
Michael A. Finegold*+ and Mathias Drton
|
|
Companies:
|
The University of Chicago and The University of Chicago
|
|
Address:
|
Department of Statistics, Chicago, IL, 60637,
|
|
Keywords:
|
Graphical Models ; Penalized Likelihoods ; Robust Inference ; EM ; Multivariate t
|
|
Abstract:
|
Gaussian graphical models have attracted a lot of recent interest. In the high-dimensional case problems can be addressed by methods of likelihood penalization. But no existing robust techniques are feasible for large problems with many variables. We propose modeling the data with a multivariate t distribution and using an L1 penalty on the elements of the concentration matrix. In the Gaussian case, a zero in the concentration matrix implies conditional independence of the corresponding variables given the rest; in the case of the t distribution a zero in the concentration matrix implies conditional uncorrelatedness. Estimation can be done using the EM algorithm. We evaluate the performance of our method on simulated data and compare our results with other methods using real data.
|
