A novel learning method for special Bayesian networks

In this paper, a novel learning method for special Bayesian networks which consist of noisy-OR and noisy-AND nodes is introduced. This method can learn networks with hidden variables and discover hidden variables when necessary. Compared with previous techniques for learning Bayesian networks, it uses the information in the data to guide the search for useful revisions, and can greatly improve the efficiency of the algorithm. Furthermore, this method can also be used for theory refinement. The experiments demonstrate that its performance is comparable to that of other existing hybrid theory refinement systems, while the networks produced by this method have more precise semantics and are more easily understood. In addition, this method also provides a direct mechanism for incorporating knowledge expressed as propositional Horn-clause rules into a Bayesian network. This mechanism could potentially ease the process of building Bayesian networks.