基于融合先验方法的贝叶斯网络结构学习

Bayesian network structures learning based on approach using incoporate priors method

从数据中学习贝叶斯网络结构是一个非确定性多项式困难(non-deterministic polynomial hard,NP-hard)问题,当数据样本不充分时难以获得准确的模型,此时利用先验信息是一种有效的途径。但是利用先验信息的过程中如何适应不正确的先验信息,是一个待解决的问题。针对此问题,提出一种融合先验的方法进行贝叶斯网络结构学习,在评分搜索法的两个环节中解决这个问题:第一,提出了新的融合不确定先验信息的评分函数,考虑了先验信息与数据集的权衡。第二,提出了融合不确定先验信息的搜索策略,增强先验信息利用的鲁棒性。所提方法适用于任何启发式搜索。仿真结果表明了所提方法能有效地利用正确的先验信息,而且对错误的先验信息有较强的适应能力。

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.