A0624
Title: Dependence structure estimation from incomplete data
Authors: Giuseppe Vinci - University of Notre Dame (United States) [presenting]
Abstract: Modern scientific multivariate data sets are often incomplete and corrupted by noise. Implementing statistical methods is very challenging in these situations, which are common in numerous science fields, including neuroscience, genomics, astronomy, and forensic science. In particular, the estimation of the covariance matrix and related objects, such as conditional dependence graphs, is nearly impossible when the data are structurally incomplete. Novel methods for the high-dimensional estimation of covariance matrices and graphical models from incomplete and corrupted data are presented. Two families of approaches are discussed: factor analysis methods and sparse precision matrix estimation methods. Theory and methods are presented for traditional random vectors and also in the framework of multivariate extreme values. Simulation studies and applications to real data are also presented.
