摘要
在引入最大信息系数的基础上,提出一种改进的贝叶斯网络结构学习算法。在给定数据集的条件下,基于最大信息系数对变量间的关联度进行检测,根据筛选因子和关联度构造贝叶斯网络的初始化结构,并结合贪婪算法对初始网络结构进行局部优化,将局部最优解进行整合形成全局最优解,生成最终的网络结构。在Asia和Car基准网络上的实验结果表明,与基于传统贪婪算法、随机K2算法的贝叶斯网络结构学习算法相比,该算法可以学习到与基准网络更相近的贝叶斯网络结构,并且具有较高的正确边均值和分类准确率。
An improved Bayesian network structure learning algorithm is proposed by introducing Maximal Information Coefficient(MIC). Under the conditions of a given data set,MIC is used to measure dependency between the variables. An initial Bayesian network is constructed according to the screening and correlation factor. It is combined with the greedy algorithm to locally modify the initial network,integrat local optimal solution to form the global optimal solution, and generate the final network structure. Experimental results on Asia and Car benchmark networks show that, compared with BN structure learning algorithm based on traditional Greedy algorithm, random K2 algorithm, the algorithm is able to get the network structure which is close to that of the benchmark network and has higher mean of the right side and classification accuracy.
出处
《计算机工程》
CAS
CSCD
北大核心
2017年第8期225-230,共6页
Computer Engineering
基金
国家自然科学基金(61300107)
广东省自然科学基金(S2012010010212)
关键词
贝叶斯网络
结构学习
最大信息系数
关联度
贪婪算法
Bayesian Network (BN)
structure learning
Maximal Information Coefficient (MIC)
relevancy
greedy algorithm