The Model Selection Methods for Sparse Biological Networks

Abstract

It is still crucial problem to estimate high dimensional graphical models and to choose the regularization parameter in dependent data. There are several classical methods such as Akaike’s information criterion and Bayesian Information criterion to solve this problem, but also more recent methods have been proposed such as stability selection and stability approach to regularization selection method (StARS) and some extensions of AIC and BIC which are more appropriate for high dimensional datasets. In this review, we give some overview about these methods and also give their consistency properties for graphical lasso. Then, we evaluate the performance of these approaches in real datasets. Finally, we propose the theoretical background of our proposal model selection criterion that is based on the KL-divergence and the bootstrapping computation, and is particularly suggested for the sparse biological networks.