具有超父结点时间序列贝叶斯网络集成回归模型  

With Super Parent Node Bayesian Network Ensemble Regression Model for Time Series

作　　者：王双成 高瑞 杜瑞杰 WANG Shuang-Cheng;GAO Rui;DU Rui-Jie(School of Information Management,Shanghai Lixin University of Accounting and Finance,Shanghai201620;School of Statistics and Mathematics,Shanghai Lixin University of Accounting and Finance,Shanghai201620)

机构地区：[1]上海立信会计金融学院信息管理学院,上海201620 [2]上海立信会计金融学院统计与数学学院,上海201620

出　　处：《计算机学报》2017年第12期2748-2761,共14页Chinese Journal of Computers

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：时间序列是现实世界中数据的主要表现形式之一,对时间序列进行预测也有着普遍的需求.现已发展了许多时间序列(单时间序列或多时间序列)预测模型,它们各有特点,被广泛用于解决诸多领域的实际问题,但这些模型或者需要时间序列平稳性和具有线性关系的假设,或者与某种分布紧密联系在一起,这使其适用范围受到限制,而且也不易于实现动态和静态信息的融合.文中在基于高斯函数估计属性密度的基础上,结合转换数据集构建、回归变量的离散化、类变量的数量化、属性联合密度的分解计算和以类的满条件概率为权重的加权平均等,建立用于时间序列预测的具有超父结点贝叶斯网络回归模型,该模型能够在统一的概率框架下实现对动态与静态信息的融合,不需要平稳性、分布和函数形式的假设,并能够通过具有不同超父结点贝叶斯网络回归模型的集成来进一步降低回归误差和提高泛化能力.使用UCI和宏观经济数据进行实验的结果显示,无论对单时间序列还是多时间序列,具有超父结点贝叶斯网络集成回归模型均具有良好的回归可靠性.Time series is one of the main forms of data in the real world.There is a wide range of demand for time series prediction.There have been many models which can be used in time series(single or multiple time series)prediction.Each of them has its characteristics and is used to solve practical problems in many fields.But these models either require time series stability and have a linear relationship between variables,or are closely associated with some kind of distribution.Their scope of application will be limited and it is also difficult to realize the fusion of dynamic and static information in these models.Bayesian network is a graph model describing dependencies between random variables and is widely used to deal with uncertainty problems.It has many features,such as multi functionality,effectiveness,openness and so on.Bayesian network can transform data into knowledge,use it to solve practical problems by reasoning and has broad application prospects in classification and regression.Although there are some researches on applying Bayesian network to regression computation,these models are essentially variants of Bayesian regression in statistics.They do not really take advantage of Bayesian networks and classification techniques,and the regression results are not satisfactory.Since the birth of the Bayesian network,it is mainly used for causal analysis(when used in causal analysis,it is also called a causal Bayesian network)and classification(when used for classification,it is called a Bayesian network classifier).Causal analysis is a classical application of Bayesian network.When Bayesian networks are used in causal analysis,variables are required to be discretized,and causal relationships between variables and quantitative reasoning are emphasized.Bayesian network classifier is built with a series of classification rules.It highlights the mapping relationship between attributes and class.The attributes of the classifier can be discrete variables or continuous and mixed variables.There is a close relation be

关 键 词：贝叶斯网络 高斯函数 时间序列 分类 回归 数据挖掘 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]