Learning Bayesian networks by learning decomposable Markov networks first
Yichao Huang, Y. Xiang · 2003
Most Bayesian network learning algorithms are based on a single-link lookahead search. The method is efficient, but it may fail when the underlying domain is complex, such as being pseudo-independent (PI). Learning PI models requires a multi-link lookahead search, which increases the complexity. Since the search space of directed acyclic graphs (DAGs) is much larger than that of chordal graphs, given the number of nodes, learning Bayesian networks directly from data using multi-link searching is very expensive. We present a Bayesian network learning algorithm which learns a decomposable Markov network as an intermediate step. Using this approach, we can learn Bayesian networks in PI domains with reduced complexity.