Traditional approaches to bayes net structure learning typically assume
little regularity in graph structure other than sparseness. However, in many
cases, we expect more systematicity: variables in real-world systems often
group into classes that predict the kinds of probabilistic dep