摘要
目前,具有已知结构的隐藏变量学习主要针对具有离散变量的贝叶斯网和具有连续变量的高斯网.该文给出了具有连续和离散变量的混合贝叶斯网络隐藏变量学习方法.该方法不需要离散化连续变量,依据专业知识或贝叶斯网络道德图中Cliques的维数发现隐藏变量的位置,基于依赖结构(星形结构或先验结构)和Gibbs抽样确定隐藏变量的值,结合扩展的MDL标准和统计方法发现隐藏变量的最优维数.实验结果表明,这种方法能够有效地进行具有已知结构的混合贝叶斯网络隐藏变量学习.
At present, the methods of learning the hidden variables of Bayesian network with known structure is mainly for Bayesian networks with discrete variables or Gaussian networks with continuous variables. In this paper, the method of learning the hidden variables of hybrid Bayesian network with discrete and continuous variables is presented. The discretization of continuous variables is not needed. The hidden variables are found by prior knowledge or the dimension of cliques in the moral graph of Bayesian network. The values of hidden variable are made based on dependency structure （star structure or prior structure） between variables and Gibbs sampling. The optimum dimension of hidden variable is made by combining extended MDL criterion with statistics method. Experimental results show that this method can effectively learn the hidden variables of hybrid Bayesian network with known structure.
出处
《计算机学报》
EI
CSCD
北大核心
2005年第9期1564-1569,共6页
Chinese Journal of Computers
基金
国家自然科学基金(60275026)
吉林省自然科学基金(200305171)
上海市重点学科项目基金(P1601)资助
