面向多层次知识表达的贝叶斯分类模型研究  

Research on Bayesian Classification Model for Multi-level Representation

作　　者：王利民[1,2] 李雄飞[1,2] 徐沛娟[1,2] 

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]吉林大学符号计算与知识工程教育部重点实验室,长春130012

出　　处：《计算机科学》2009年第3期119-122,共4页Computer Science

基　　金：国家自然科学基金(60275026)资助

摘　　要：提出多模式贝叶斯分类算法,由变量值之间的条件独立和条件相关性推断因果关系,根据每个完整随机样本而非整个样本空间构造子模式。结合局部计算近似推理进行概率密度和条件概率分布估计,在此基础上采用后离散化策略自动确定连续变量边界。在UCI机器学习数据集上的实验结果证明了该算法的合理性和有效性。A multi-schema Bayesian classification algorithm was proposed to solve the problem of discretization assumption and graph representation. By reasoning the conditional independence and dependence between attribute values, a submodel was constructed for each complete random instance rather that the whole instance space. The boundary of continuous attribute was decided automatically based on post-discretization strategy, the joint probability density arid conditional probability were estimated based on marginal computation. The experimental study on the UCI data set shows that, this algorithm overcomes the restrictiveness of traditional TAN and can describe the marginal dependency of mixed-mode data more intuitively and accurately.

关 键 词：贝叶斯网络 多模式 后离散化策略 局部计算 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] O212.1[自动化与计算机技术—控制科学与工程]