马尔科夫毯学习算法综述.doc

马尔科夫毯学习算法研究和进展* 傅顺开，Sein Minn (华侨大学 计算机科学与技术学院,福建省厦门市 361021） 摘 要: 马尔科夫毯（MB）在贝叶斯网络（BN）研究中较早被认识和定义，它是BN拓扑结构的重要组成部分。在1996年被证明是预测目标变量的最优特征子集。给定全局BN可以很容易推导出特定变量的MB，但BN结构的学习已知是NP问题。回顾了从1996年至今关于MB学习算法的17个典型工作，包括 (1) 基于MB的全局条件独立特征的KS、IAMB等8个算法，(2) 利用MB的局部条件独立特征的PCMB、IPC-MB等6个算法, （3）基于逻辑回归分析+局部条件独体特征的RA-MMMB算法，和 (4) 基于评分-搜索方法的DMB和RPDMB两个算法。讨论时兼顾理论和实用性内容，并统一/扩展了相关算法的伪代码（描述），对学术界和工业界研究人员都具参考价值。A Review of Markov Blanket Induction Algorithms FU Shun-kai, SEIN Minn (College of Computer Science and Technology, Huaqiao University, Xiamen Fujian 361021, China) Abstract: Markov blanket (MB) has been realized and defined during the research of Bayesian network, and it is an important component of the BN. In 1996, it was proved the optimal feature subset for prediction. Given the global BN, it is trivial to read off the MB of specific variable, but the structure learning of BN is known as NP-hard problem. From 1996 on, there are many published works on the induction of MB, and our review covers 17 typical works, including (1) those based on the global conditional independence (CI) probability feature of MB, like KS, IAMB etc., (2) those based on the local CI probability feature, such as PCMB, IPC-MB etc., (3) one work with the combination of logistic regression and constraint learning, called RA-MMMB, and (4) non-constraint learning works based on score-and-search, DMB and RPDMB. Our discussion covers theoretical as well as practical aspect, and we revise the pseudo codes of related algorithms to easier the understanding. It is believed a useful reference for both academic and industrial colleagues. Key words: Markov blanket;Bayesian network; constraint learning;dimension reduction; feature selection 引言 贝叶斯网络（Bayesian Network, BN）是一种有向无环图（DAG）模型，其中节点代表了随机变量，而边代表了随机变量之间的概率关系。基于条件独立性，BN的图模型能够有效紧凑表达目标问题的联合概率关系，并可以通过贝叶斯链式法则来快速实现从图表征语言到公式的相互转换。BN是人工智能领域的一种重要工具，被成功运用到机器学习和数据挖掘领域。 早在1988年，Peal就在他的关于贝叶斯网络研究的专著[1]里定义和讨论了马尔科夫毯（Markov Blanket，MB）。给定一个节点，它的MB是唯一的，包括的所有父、子和配偶