基于完全代理模型的贝叶斯网络结构学习  

Bayesian network structure learning based on complete surrogate model

作　　者：王慧玲[1,2] 綦小龙 梁义[1] WANG Huiling;QI Xiaolong;LIANG Yi(Department of Electronics and Information Engineering,Yili Normal University,Xingjiang Yining 835000,China;Department of Computer Science and Technology,Nanjin University,Nanjin 210023,China)

机构地区：[1]伊犁师范学院电子与信息工程分院,新疆伊宁835000 [2]南京大学计算机科学与技术系,南京210023

出　　处：《激光杂志》2018年第9期128-132,共5页Laser Journal

基　　金：国家自然科学基金项目(No.61663045,No.61761043);伊犁师范学院校级科研项目(No.2015YSYB30)

摘　　要：贝叶斯网络结构学习中,基于变量序的空间搜索方法不仅是高效的而且结构的质量也是可靠的。但是,在搜索过程中,序的质量评估是一个非常关键的而且也是困难的问题。现有的方法虽然是高效的,但是序的质量评估不可靠。本文提出了一种新的评估序的质量的方法即完全代理模型评估法,该模型用近似于真实的父集来确定变量序的质量,从而找到一个高评分的网络结构。提议的方法主要包含两部分：最佳邻居学习,在变量数的多项式时间内使用互信息学习每个变量的最佳邻居集合;最佳父集学习,在最佳邻居集合规模的指数级时间内,根据当前的变量顺序以及变量的邻居集合学习每个变量的最佳的父集。本文从理论上分析了算法的合理性,从实验上和现有的代理模型算法做了对比,验证了算法的可靠性。In Bayesian network structure learning,the space search method based on variable order is not only efficient but also the quality of the structure is reliable. However,the quality evaluation of the order is a very critical and difficult problem in the search process. Although the existing method is efficient,the quality evaluation of the order is not reliable. This paper proposes a new method to evaluate the quality of the order,that is,an evaluation method of complete surrogate model. This model uses approximation to the real parent set to determine the quality of the variable order and finds a high score network structure. The proposed method mainly consists of two parts： the best neighbor learning,uses mutual information to learn the best neighbor set for each variable,and the best parent set learning,learns the best parent of each variable based on the current variable order and the neighbor set of variables. This paper theoretically analyzes the rationality of the algorithm and compares it experimentally with the existing surrogate model algorithm to verify the reliability of the algorithm.

关 键 词：贝叶斯网络结构 代理模型 变量序 

分 类 号：TN181[电子电信—物理电子学]