A Variational Approach for Approximating Bayesian Networks by Edge Deletion