A0172
Title: Sparsity-constrained estimators for graphical models
Authors: Alessandro Fulci - University of Trento (Italy) [presenting]
Abstract: Graphical models provide a versatile framework for representing conditional dependence structures among random variables. New methods are proposed for estimating a sparse and shrunk precision matrix in undirected Gaussian graphical models. New approaches are introduced that, besides incorporating the l1-norm penalty, rely on a constraint on the l0-pseudo-norm. Specifically, the newly introduced estimators include the sparsity-constrained Glasso (SCGlasso), the adaptive graphical Lasso (AGlasso), and its sparsity-constrained variant (SCAGlasso). An essential feature of our proposed algorithms is their ability to circumvent the highly non-convex nature associated with the l0-constraint. A comprehensive comparison with established methods like Glasso and the Atan penalty is conducted through simulations, revealing the effectiveness of the l0-constraint, particularly in model selection and for small sample sizes. Additionally, a real-world application is explored in the context of gene expression, supporting the validity of the proposed approaches.
