约束高斯分类网研究  

Restricted Gaussian Classification Network

作　　者：王双成[1,2] 高瑞[1,3] 杜瑞杰[1] 

机构地区：[1]上海立信会计学院数学与信息学院,上海201620 [2]上海立信会计学院立信会计研究院,上海201620 [3]上海财经大学统计与管理学院,上海200433

出　　处：《自动化学报》2015年第12期2164-2176,共13页Acta Automatica Sinica

基　　金：国家自然科学基金(61272209);上海市自然科学基金(15ZR1429700);上海市教委科研创新项目(15ZZ099)资助~~

摘　　要：针对基于一元高斯函数估计属性边缘密度的朴素贝叶斯分类器不能有效利用属性之间的依赖信息和使用多元高斯函数估计属性联合密度的完全贝叶斯分类器易于导致对数据的过度拟合而且高阶协方差矩阵的计算也非常困难等情况,在建立属性联合密度分解与组合定理和属性条件密度计算定理的基础上,将朴素贝叶斯分类器的属性选择、分类准确性标准和属性父结点的贪婪选择相结合,进行约束高斯分类网学习与优化,并依据贝叶斯网络理论,对贝叶斯衍生分类器中属性为类提供的信息构成进行分析.使用UCI数据库中连续属性分类数据进行实验,结果显示,经过优化的约束高斯分类网具有良好的分类准确性.Naive Bayes classifier using unitary Gaussian function to estimate attribute marginal density can not effectively use dependency information between attributes, and the full Bayes classifier using multivariate Gaussian function to estimate attribute joint density often leads to over fitting and the difficulty to calculate high-order covariance matrix. In this paper, based on Gaussian network theory, a decomposition and combination theorem for attribute joint density and a calculation theorem for attribute conditional density are established. A restricted Gaussian classification network is presented by combining the attribute selection of naive Bayes classifier, the evaluation criteria of classification accuracy and the greedy search of attribute parent nodes. The information composition of attributes provided for class is analyzed in Bayesian derivative classifiers. Experiment and analysis are done by using data sets with continuous attributes in UCI.The results show that restricted Gaussian classification networks have very good classification accuracy.

关 键 词：连续属性 高斯网络 贝叶斯网络 贝叶斯衍生分类器 密度估计 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]