Title: Constraint-Based Learning Bayesian Networks Using Bayes Factor
Abstract: A score-based learning Bayesian networks, which seeks the best structure with a score function, incurs heavy computational costs. However, a constraint-based (CB) approach relaxes this problem and extends the available learning network size. A severe problem of the CB approach is its lower accuracy of learning than that of a score-based approach. Recently, several CI tests with consistency have been proposed. The main proposal of this study is to apply the CI tests to CB learning Bayesian networks. This method allows learning larger Bayesian networks than the score based approach does. Based on Bayesian theory, this paper addresses a CI test with consistency using Bayes factor. The result shows that Bayes factor with Jeffreys’ prior provides theoretically and empirically best performance.