以核心变量为基础的离散贝叶斯网络结构学习  

Learning discrete Bayesian network structures from data:based on kernel variables

作　　者：张邦佐[1] 王辉[1] 张剑飞[1] 左万利[2] 

机构地区：[1]东北师范大学计算机学院,吉林长春130024 [2]吉林大学计算机科学与技术学院,吉林长春130025

出　　处：《东北师大学报（自然科学版）》2005年第4期28-31,共4页Journal of Northeast Normal University(Natural Science Edition)

基　　金：吉林省自然科学基金资助项目(20030517-1);东北师范大学青年教师基金资助项目

摘　　要：建立了基于核心变量的离散贝叶斯网络结构学习方法.该方法根据变量之间的无条件相对预测能力建立有向无环图,分别按着变量的聚度和散度排序变量;以不同于被预测变量的具有最大聚度和散度的两个变量为条件变量,根据变量之间条件相对预测能力的大小确定弧的存在性与方向,结合环路检验建立初始贝叶斯网络结构;以两个变量的最小切割集为条件变量集,调整初始贝叶斯网络结构(包括删除多余的弧和重新确定弧的方向),最终建立数据中所蕴涵的贝叶斯网络结构.同时,使用模拟数据进行了对比实验,结果表明这是一种有效实用的方法.In this paper, based on kernel variables, the method of learning discrete Bayesian network structures from data was developed. This method is made up of three parts. First, a directed acyclic graph is built in terms of unconditional relative forecasting ability between variables and the variables are sorted degressively according to the convergence degree and divergence degree of variables. Second, two variables, which respectively have maximum convergence degree and divergence degree and are different from forecasted variables, are selected as conditional variables. The existence and direction of arc between two variables are made in terms of conditional relative forecasting ability and an elementary Bayesian network structure is built with checking cyclic route. Third, given conditional set（minimum d-separating set of two variables）, the elementary Bayesian network structure is regulated （to increase the lost arcs, to delete superfluous arce and to regulate direction of arcs）in terms of conditional relative forecasting ability and a Bayesian network structure is built with checking cyclic route. In the mean time, a contrast experiment was made by using simulated data.

关 键 词：预测能力 核心变量 最小切割集 聚度 散度 

分 类 号：TP311.131[自动化与计算机技术—计算机软件与理论]