异步动态贝叶斯网络分类器研究  

Asynchronous dynamic Bayesian network classifiers

作　　者：王双成 张立 郑飞 WANG Shuang-Cheng;ZHANG Li;ZHENG Fei(School of information management,Shanghai Lixin University of Accounting and Finance,Shanghai 201620;Institute of data science and interdisciplinary studies,Shanghai Lixin University of Accounting and Finance,Shanghai 201209)

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数据科学交叉研究院,上海201209

出　　处：《计算机学报》2020年第9期1737-1754,共18页Chinese Journal of Computers

基　　金：国家社会科学基金(18BTJ020)资助.

摘　　要：时间序列数据普遍存在,对其进行分类预测有着广泛的需求.虽然有一些时间序列数据分类方面的研究,但主要是面向时序同步分类(类与属性同步变化),还需要进行更有实际意义的异步分类(类与属性不同步变化)方面的探索.本文结合时间序列的离散化、变量的时序转换、变量的错位变换、依据变量顺序和打分搜索的分类器结构学习和类变量的丢失数据修复等,建立异步动态贝叶斯网络分类器,这种分类器能够有效利用多变量时间序列数据中所蕴含的时滞、非时滞和混合分类信息,以及属性为类提供的传递依赖信息、直接导出依赖信息和间接导出依赖信息进行分类计算,来提高分类器的可靠性.分别使用UCI、金融和宏观经济时间序列数据进行实验,结果显示所建立的异步动态贝叶斯网络分类器具有良好的分类准确性.Time series is one of the main forms of data in the real world.It widely exists in various large databases,such as macroeconomic,finance,industry,management,internet and so on.A large number of time series record all kinds of important information of the system at different time points(or time slices).There is abundant and valuable knowledge about causality,classification rules and regression functions in these information.They are often the important basis for diagnosing the operation of the system and formulating corresponding strategies.Classification is a computer simulation of human concept learning(also known as concept learning).It is one of the core technologies in machine learning and data mining.Many famous classifiers have been developed,such as decision tree,neural network,support vector machine,nearest neighbor classifier and so on.They have their own characteristics and are widely used in many fields,but these classifiers are mainly for non time series data.Bayesian network is a probabilistic graphical model to describe the dependence and restriction relationship between random variables.It has the characteristics of multi-function,effectiveness and openness,and is a powerful tool to deal with uncertainty.Classical Bayesian networks are mainly used for causal knowledge representation and uncertainty reasoning.The Bayesian network for classification is generally called Bayesian network classifier.There are many researches on Bayesian network classifiers,but these classifiers are all for non time series data and can not be directly used in the classification calculation of time series data.Dynamic Bayesian network is a timing extension of Bayesian network.It is mainly used to solve the uncertainty of time series.The dynamic Bayesian network for time series data classification is generally called dynamic Bayesian network classifier.The research on dynamic Bayesian network classifier is relatively less,and it is mainly synchronous classification(synchronous change of class and attributes).It's also need

关 键 词：时间序列 动态贝叶斯网络 分类器 同步分类 异步分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]