基于因果效应的贝叶斯网络结构学习方法  

Bayesian network structure learning method based on causal effect

作　　者：安宁[1,2] 滕越 杨矫云 李廉[1,2] An Ning;Teng Yu(National Smart Eldercare International S＆T Cooperation Base;School of Computer ＆ Information,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]合肥工业大学国家智慧养老国际科技合作基地,合肥230009 [2]合肥工业大学计算机与信息学院,合肥230009

出　　处：《计算机应用研究》2018年第12期3609-3613,共5页Application Research of Computers

基　　金：安徽省重点研究与开发计划资助项目(1704e1002221);国家自然科学基金青年基金资助项目(61502135);国家高等学校学科创新引智计划资助项目(B14025)

摘　　要：从数据中学习贝叶斯网络结构是一个NP-hard问题,提高网络结构学习算法精度是研究的重难点。基于Pearl的因果理论,提出了一种贝叶斯网络结构学习方法,提升了现有算法的准确率。利用改进的Pearl因果效应和BDe评分,学习网络节点优先次序,利用K2算法学习初始网络,并通过BDe评分反向调节、互信息和BDe评分删除边以修正学习结果。实验在贝叶斯网络标准数据集ASIA、ALARM上进行,在样本量为2000～20 000的20组实验中,学习准确率较MMHC算法平均提升16%,准确率标准差较MMHC算法平均缩小17%。实验表明,基于因果效应的方法较MMHC算法有更好的性能。Learning Bayesian networks from data is a NP-hard problem. Improving the accuracy of network structure learning algorithms is important,yet difficulty for researchers. The paper proposed a Bayesian network structure learning algorithm based on Pearl＇s causal theory to improve the accuracy of existing algorithms. The algorithm made use of the improved causal effect and the BDe scoring-function to learn the nodes order. Then applied K2 algorithm to achieve an initialized network. At last,it used the BDe scoring-function to adjust the network inversely,besides,it used mutual-information and BDe scores to prune the edges. The proposed method outperforms MMHC algorithm on the ASIA and ALARM data set,with 16% accurate improvement,and 17% standard deviation less on average. The results show that the method based on causal effect is batter then MMHC algorithm.

关 键 词：贝叶斯网络 阿尔茨海默病 K2算法 因果效应 BDe评分 互信息 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]