Towards Effective Structure Learning for Large Bayesian Networks

The effectiveness of Bayesian network construction is a function of the predictive ability and speed of inference of the resulting network, the effort and expertise required to construct the network, and the effort and expertise required to understand the network. We empirically evaluate three alternative methods for constructing Bayesian networks from data, considering both objective performance measures and subjective construction cost measures. Empirical results are obtained for a large real-world medical database. We provide results comparing network structure laboriously elicited from a domain expert to structure automatically induced by two alternative structure learning algorithms. The parameters of the Bayesian network produced by each method are induced using the Bayesian MAP approach. Additionally, we introduce the use of classification paths as an aggregation technique to reduce the size and structural complexity of Bayesian networks. Empirical results are obtained both with and without this complexity reduction technique.