Capitalizing on asymmetric relationships when learning the structure of Bayesian belief networks
Carol A. Wellington · 1997
While there has been recent work on efficiently encoding asymmetric relationships in Bayesian belief networks (9, 14), relatively little research (2) directly addresses how asymmetric relationships affect structure learning algorithms. This thesis demonstrates that the asymmetry of relationships can affect structure learning algorithms, and defines a framework which will allow any structure learning algorithm to be more sensitive to asymmetric relationships. The framework operates by first learning a similarity network from various subsets of the available training data and then builds a single Bayesian network from the similarity network. The value of this learning framework is demonstrated by using it in combination with minimum description length (MDL) and Advanced Pattern Recognition & Identification (APRI) structure learning algorithms. When modeling populations exhibiting asymmetric relationships, both learning algorithms show a dramatic run-time benefit without compromising the accuracy of the resulting network. In addition, using the framework with APRI may result in improved model accuracy and inference run-times.