摘要
为了得到正确的节点次序,构造接近最优的贝叶斯网络结构,利用最大信息系数与条件独立性测试相结合的方法,提出了一种新的贝叶斯网络结构学习算法(MICVO)。该算法利用最大信息系数衡量变量之间的依赖关系,生成初始的无向图,引入惩罚因子δ减少图中冗余边的数量,并将这个无向图分解成多个子结构,确定图中边的方向,最后生成正确的节点次序作为K2算法的输入学习网络结构。在两个基准网络Asia和Alarm中进行实验验证,结果表明基于最大信息系数的贝叶斯网络结构学习算法可以得到接近最优的节点次序,学习到的网络结构与数据的拟合程度更好,分类准确性更高。
In order to obtain the correct node ordering,this paper presented a new Bayesian network structure learning algorithm （MICVO）which used the method based on maximal information coefficient combining with conditional independence test.First-ly,it generated an initial undirected graph through measuring dependency between variables using maximal information coeffi-cient,and introduced a penalty factorδto reduce the number of redundant edges.Then divided this undirected graph into multi-ple sub-structures to determine the direction of edges in the graph,and finally the initial ordering of nodes obtained was as input of K2 algorithm to construct the network structure.Experimental results over two benchmark networks Asia and Alarm prove that the Bayesian network structure learning algorithm based on maximal information coefficient can obtain bear optimal ordering of nodes,network structure with better degree of data matching,and higher classification accuracy.
出处
《计算机应用研究》
CSCD
北大核心
2014年第11期3261-3265,共5页
Application Research of Computers
基金
国家自然科学基金资助项目(61172090)
关键词
贝叶斯网络
结构学习
节点次序
最大信息系数
条件独立性测试
Bayesian network（BN）
structure learning
node ordering
maximal information coefficient（MIC）
conditional in-dependence test
