The Closed V-Structure Approach to Bayesian Network Learning

A new heuristic for learning the structure of belief networks is introduced. In this closed v-structure approach, the space of node orderings is searched, with one ordering preferred to another whenever its greedily chosen structure has fewer closed v-structures. A closed v-structure is a configuration of nodes wherein two parents of some node have a direct dependency relation. The number of arcs in a network is generally reduced by reversing the ordering of nodes involved in such structures, and this fact is taken advantage of here. This approach proves to be very effective at rapidly learning a good network structure, although it does not necessarily learn the optimal structure.