Learning Bayesian Networks with Incomplete Data by Augmentation