We present a novel structural learning method called HdBCS that performs covariance selection in a Bayesian framework for datasets with tens of thousands of variables. HdBCS is based on the intrinsic connection between graphical models on undirected graphs and graphical models on directed acyclic graphs (Bayesian networks). We show how to produce and explore the corresponding association networks by Bayesian model averaging across the models identified. We illustrate the use of HdBCS with an example from a large-scale gene expression study of breast cancer.

Keywords: Bayesian inference; Bayesian model averaging; Bayesian networks; Covariance selection; Directed acyclic graphs; Gene expression data; Graphical models.

The manuscript is available in PostScript and PDF formats.