Bayesian Network Learning with Abstraction Hierarchies and Context-Specific Independence