Abstract. Copula Bayesian networks are powerful probabilistic mod- els to represent multivariate continuous distributions, while comes along with a simplified parameter estimation with flexible choices of univariate marginals. In particular form of Gaussian copula, model parameters can be efficiently estimated by inverse normal transformation. Nevertheless, learning the underlying causal graph remains problematic. In this paper, we propose a method for structure learning of copula Bayesian networks based on estimated precision matrix for copulas, named PreCBn. We show that the method gives good estimation of the networks and outperforms other related approaches of structure learning. The results indicate that the PreCBn is able to recover known causal relations more efficiently than existing approaches, and consequently results in graphical models with accurate structure and simplified model parameters.
see experiments.py