Decomposing Local Probability Distributions in Bayesian Networks for Improved Inference and Parameter Learning