Bayesian Network Structure Learning Algorithm Based on Score Increment and Reduction

Most score-based approaches of the Bayesian networks typically employ greedy search strategies, which optimize the local structure unconsciously and get stuck into the local optimum easily. Inspired by the decomposability of scoring function, this paper proposes a structure learning algorithm based on score increment and reduction. Firstly, the edge with the highest score increment is added under the guidance of the profit table. Because the previous operation ignores the acyclic constraint, it is necessary for some strategies, such as depth-first search to find all cycles. Then, the current structure should be thinned by deleting edges and clearing cycles on the basis of the loss table with score reduction. The optimal structure is acquired by repeating the above search process until the profit table is empty. Experiments show that the proposed algorithm has better performance of scoring results and graphical accuracy than some state-of-the-art structure learning algorithms in seven networks with different sample sizes.