[2D6-GS-2] Machine learning: Bayesian estimation

[2D6-GS-2-02] Learning large Bayesian networks having more than ten thousand variables

〇Koya Kato¹, Maomi Ueno¹ (1. The University of Electro-Communications)

Keywords:Bayesian network, Probabilistic graphical model, Bayes

Recent reports of some studies have described that constraint-based learning methods using the Bayes factor effectively learn large Bayesian networks to approximate a joint probability distribution. These methods can learn large networks, which previous methods are unable to do. However, learning accuracies of these methods become worse because the reliability of the Bayes factor become worse of learning large Bayesian networks. To resolve this problem, this study proposes a new algorithm using the highly reliable Bayes factor. Specifically, the proposed method (1) learns large Bayesian networks using the highly reliable Bayes factor and (2) conducts the exact learning given the structure learned in (1). The proposed method improves the learning accuracies. Comparison experiments demonstrated that the proposed method can learn a Bayesian network with more than ten thousand variables accurately.