Learning BBN Structure

• Bayesian scoring metric for choosing among alternative networks.
• The K2 algorithm performs greedy search that trades off network complexity for accuracy over the training data when the data is fully observable.
• In experiments K2 was given 3,000 training examples generated at random from a manually constructed BBN containing 27 nodes and 46 arcs representing anesthesia problems in hospital operating rooms. In addition it was given a initial ordering over the 37 variables that was consistent with the partial ordering in the actual BBN. K2 reconstructed the BBN structure with only the lost of one arc and the addition of another.
• Other approaches infer dependence and independence relationships from the data and then use these relationships to construct BBN.
• This is a very active area of research.