计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于最大信息系数的贝叶斯网络结构学习算法

曾千千 1,曾安 1,潘丹 2,杨海东 3,邓杰航 1   

  1. (1.广东工业大学 计算机学院,广州 510006; 2.广州市本真网络科技有限公司,广州 510665; 3.中油管道物资装备总公司,河北 廊坊 065000)
  • 收稿日期:2016-05-05 出版日期:2017-08-15 发布日期:2017-08-15
  • 作者简介:曾千千(1992—),女,硕士研究生,主研方向为贝叶斯网络、数据挖掘;曾安(通信作者),教授、博士;潘丹,高级工程师、博士;杨海东,工程师、学士;邓杰航,副教授、博士。
  • 基金项目:
    国家自然科学基金(61300107);广东省自然科学基金(S2012010010212)。

Bayesian Network Structure Learning Algorithm Based on Maximal Information Coefficient

ZENG Qianqian 1,ZENG An 1,PAN Dan 2,YANG Haidong 3,DENG Jiehang 1   

  1. (1.School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China; 2.Guangzhou Benzhen Network Technology Co.,Ltd.,Guangzhou 510665,China; 3.China Petroleum Pipeline Material and Equipment Corporation,Langfang,Hebei 065000,China)
  • Received:2016-05-05 Online:2017-08-15 Published:2017-08-15

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。

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

Abstract: 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.

Key words: Bayesian Network(BN), structure learning, Maximal Information Coefficient(MIC), relevancy, greedy algorithm

中图分类号: