多维贝叶斯分类器的分析-计算机软件与理论专业论文.docxVIP

摘 要 多维贝叶斯分类模型是一种描述类变量与类变量、类变量与属性变量和属性 变量与属性变量之间的依赖关系的分类模型．本文针对完全朴素多维贝叶斯分类 模型要求变量之间具有独立性假设，以及无法有效利用类变量和类变量，属性变 量与属性变量之间的依赖关系等问题，研究了多维贝叶斯分类模型并进行了仿真 实验． 首先基于贝叶斯网络的最大似然值的打分方法，学习多维贝叶斯分类模型的 类子图，属性子图和桥子图．再通过加边、减边或者转边等措施对最初的结构图 进行改进，得到一个分类效果较好的多维贝叶斯分类模型． 其次针对数据包含影响分类效果的冗余属性和无关属性，本文提出了一种新 的基于马尔科夫毯的属性选择方法：把分类变量作为普通节点，学习贝叶斯网络 结构，在此结构中寻找类变量的子节点和配偶节点，进一步在给定类变量时寻找 与类变量的马尔科夫毯节点集条件独立的变量集，进而删除冗余属性节点． 最后对三类数据进行仿真实验，结果表明本文采用的属性选择方法能够有效 删除冗余属性变量，分类效果优于对比的其它分类算法． 关键词：多维贝叶斯分类器; 最大似然打分; 属性选择; 马尔科夫毯 Abstract The multi-dimensional Bayesian classification model is a model which describes relationship of class variables and class variables, class variables and feature variables, feature variables and feature variables ． The fully na?ve multi-dimensional Bayesian classification model is based on the assumption that variables are independent and it can not capture the relationship of class variables and class variables, feature variables and feature variables ． For the above problems, we research on the multi-dimensional Bayesian classification model and do the simulation experiment． Firstly, based on the method of the maximum log-likelihood scoring, We respectively learn the class subgraph, feature subgraph and bridge subgraph by different scoring function ． Then we make the initial structure better by edge addition, edge deletion and edge reversal to get the improved multi-dimensional classification model whose classification result can be improved greatly． Secondly, we put forward a new attribute selection method which is based on the Markov blanket for the redundant attributes and irrelevant attributes which are included in some data sets effect the classification result．We learn the ordinary Bayesian network structure by organizing the class variables as the ordinary variables．Then find out the Markov blanket set of the class variables and the set which is conditional independent from the Markov blanket set ． Accordingly, the other redundant feature variables are canceled． Finally, we did