Structure Learning for Gaussian Graphical Models

By Benjamin Marlin

Estimating a single multivariate Gaussian distribution is a challenging problem when the number of data cases N is less than (or even close to) the number of data dimensions D. This is due to the fact that the standard maximum likelihood estimate for the covariance matrix is invalid when N<D, and is highly sensitive when N>D but D/N is close to one. One solution to dealing with this problem is to regularize the Gaussian distribution by requiring the inverse covariance matrix to be sparse. This corresponds to learning a Gaussian Graphical Model (GGM) with sparse connectivity structure. In this talk I will give an overview of our recent work on structure learning for GGMs including our work on accelerating stochastic local search for GGMs (to appear at NIPS'09), and our work on learning sparse GGMs with latent block structure (ICML'09 and UAI'09).

This is joint work with Kevin Murphy, Mark Schmidt, and Emtiyaz Khan at UBC and Baback Moghaddam at NASA/JPL.

Visit the LCI Forum page