一种利用不可行解的贝叶斯网学习算法

An Approach to Learning Bayesian Network by Using Infeasible Solutions

摘要:

现有的基于打分搜索的贝叶斯网学习方法都是利用满足有向无环图的可行解进行学习。在搜索过程中遇到不可行解时，这类算法简单地去除不可行解或将不可行解转化为可行解。然而，有的不可行解中往往蕴含着有价值的信息。文中提出一种新的贝叶斯网学习方法ISEC，该方法同时利用可行解和不可行解学习贝叶斯网络，并提出针对不可行解的选择策略，可以在学习过程中有效地利用不可行解中的有用信息。实验结果表明，ISEC能够比仅利用可行解的方法更快地学习到更优的贝叶斯网。

Abstract:

Existing Bayesian network learning approaches based on search & scoring usually work with feasible solutions which satisfy directed acyclic graph. This kind of approaches often removes infeasible solutions or converts infeasible solutions to feasible solutions when the solutions are infeasible. However, some infeasible solutions maybe have more useful information. This paper proposes the ISEC method for learning Bayesian network by using feasible and infeasible solutions synchronously based on an infeasible solution selection strategy. Then, the method can take advantage of the information in the infeasible solutions. Experiments show that the proposed approach can achieve better performance in less time than the approaches which only use feasible solutions.