基于改进BIC评分的贝叶斯网络结构学习

引入专家知识已成为小数据集条件下贝叶斯网络建模的主流方法,然而,专家知识是否正确直接决定了算法的结果和性能。因此,在考虑专家知识正确性的基础上,本文对贝叶斯网络结构学习问题展开研究。首先,建立一种基于连接概率分布的结构约束模型来表示专家知识,进而结合该约束模型对贝叶斯信息准则(Bayesian information criterions,BIC)评分进行改进;最后,利用K2算法学习贝叶斯网络结构。实验结果表明,在小数据集条件下本文所提算法不仅能将专家知识引入到学习过程中,进而改善学习效果,并且对不完全正确的专家知识有一定的适应性。

Bayesian networks structure learning based on improved BIC scoring

Introducing expert knowledge is the main method of Bayesian networks(BN) modeling from small data set. The results and performance of algorithm are affected by the correctness of the expert knowledge. Therefore, considering the correctness of the expert knowledge, the problem of BN learning is studied. First of all, the structural constraints model based on joint probability distribution is proposed to represent the expert knowledge, and then the Bayesian information criterions (BIC) is improved by combining with the constraint model. Finally, the K2 algorithm is used for learning BN. The experimental results show that the proposed algorithm can not only introduce the expert knowledge into the process of BN learning to improve the learing effect, but also have some adaptability to the not entirely correct expert knowledge.