基于融合先验方法的贝叶斯网络结构学习

Learning Bayesian network structures from data is an non-deterministic polynomial hard pro-blem. It is difficult to get an accurate model when the data is sparse, at this point, using prior knowledge is a valid approach. However, it is an unsolved problem that how to deal with incorrect prior knowledge in the process of using it. To solve this problem, an approach using priors to learn Bayesian network structures is proposed and this problem is soloved in two phases of search and score algorithms. First, a score function is proposed which incorporates uncertain prior knowledge and the trade-off between prior knowledge and training data is considered. Second, a search strategy that incorporates uncertain prior knowledge is proposed, which strengthens the robustness of using priors. Besides, this strategy is suitable for any heuristic search process. Simulation results show that the proposed methods can effectively utilize the correct prior knowledge, and have certain adaptability for some wrong priors.