正在加载图片...
第7卷第2期 智能系统学报 Vol.76.2 2012年4月 CAAI Transactions on Intelligent Systems Apr.2012 D0I:10.3969/i.issn.16734785.201112005 网络出版t地址:htp://www.cnki.net/kcma/detail/23.1538.TP.20120416.0843.001.html 逻辑回归分析的马尔可夫毯学习算法 郭坤,王浩,姚宏亮,李俊照 (合肥工业大学计算机与信息学院,安微合肥230009) 摘要:针对当前的马尔可夫毯学习算法会引入不正确的父子节点和配偶节点的问题,提出了一种基于逻辑回归分 析的马尔可夫毯学习算法RA-MMMB.利用MMMB算法得到候选的马尔可夫毯,建立目标变量与候选马尔可夫毯的 逻辑回归方程,通过回归分析在保留与目标变量相关性很强的变量的同时,去掉MMB等算法所引入的弱相关性的 错误变量以及其他的弱相关性变量;然后利用G测试去掉回归分析后候选马尔可夫毯中的兄弟节点,得到目标变 量的马尔可夫毯.RA-MMMB算法通过回归分析,减少了条件独立测试的次数,提高了学习的精度.实验比较和分析 表明,RA-MMMB算法能有效地发现变量的马尔可夫毯. 关键词:贝叶斯网络;马尔可夫毯:逻辑回归分析:条件独立测试 中图分类号:TP181文献标志码:A文章编号:16734785(2012)02-015308 An algorithm for a Markov blanket based on logistic regression analysis GUO Kun,WANG Hao,YAO Hongliang,LI Junzhao College of Computer and Information,Hefei University of Technology,Hefei 230009,China) Abstract:To solve the problem of incorrect parent,child,and spouse nodes being brought into the current algo- rithms,an improved algorithm called a regression analysis-max min Markov blanket(RA-MMMB)was presented u- sing the Markov Blanket based on logistic regression analysis.First,a logistic regression equation was established between the target variable and a set of its candidate Markov blankets obtained from the max-min Markov blanket (MMMB)algorithm.Regression analysis can retain the variables strongly correlated with the target variable,and can remove the error variables and other variables weakly correlated with it as well.The incorrect nodes in the MMMB algorithm were also removed from the candidate Markov blanket;then,after the G2 conditiond independ- ence test,which removed the brother node of the target variable in the candidate Markov blanket,returned after the regression analysis,the Markov blanket of the target variable was obtained.By the method of regression analysis, the RA-MMMB algorithm reduces the number of condition tests of independence and improves the accuracy of dis- covering the Markov blanket for the target variable.The result shows that the method can discover the Markov blan- ket of the target variable efficiently. Keywords:Bayesian networks;Markov blanket;logistic regression analysis;conditional independence test 在给定贝叶斯网络(Bayesian networks)中一个打分一搜索方法等建立贝叶斯网络结构,然后基于 变量的马尔可夫毯(Markov blanket)时,贝叶斯网络 贝叶斯网络结构确定目标变量的马尔可夫毯,但该 中其他变量与该变量条件独立,一个变量的马尔可 类方法得到的马尔可夫毯不准确,且学习方法效率 夫毯能够屏蔽贝叶斯网络中其他变量对该变量的影 低:另一类是利用局部学习的方法直接学习目标变 响,可用来预测、分类和因果发现等 量的马尔可夫毯.当前研究者主要采用基于局部学 确定目标变量的马尔可夫毯有2类方法:利用 习的方法学习马尔可夫毯,相关工作如Margaritis和 Thrun提出了GS(Gow-Shrink)算法I,首先启发式 收稿日期:2011-1208.,网络出版日期:201204-160. 地搜索所有与目标变量依赖的变量,然后去除冗余 基金项目:国家自然科学基金资助项目(61070131,61175051) 通信作者:郭坤.E-mail:guokun19871005@163.com. 的变量.由于配偶节点较晚进入候选的马尔可夫毯
向下翻页>>
©2008-现在 cucdc.com 高等教育资讯网 版权所有