Page 96 - 《软件学报》2021年第10期

软件学报 ISSN 1000-9825, CODEN RUXUEW E-mail: jos@iscas.ac.cn
Journal of Software,2021,32(10):30683084 [doi: 10.13328/j.cnki.jos.006012] http://www.jos.org.cn
©中国科学院软件研究所版权所有. Tel: +86-10-62562563


基于贝叶斯网络的时间序列因果关系学习 

1
1,2
王双成 , 郑 飞 , 张 立 1
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)
通讯作者: 王双成, E-mail: wangsc@lixin.edu.cn

摘 要: 贝叶斯网络是研究变量之间因果关系的有力工具,基于贝叶斯网络的因果关系学习包括结构学习与参数
学习两部分,其中,结构学习是核心.目前,贝叶斯网络主要用于发现非时间序列数据中所蕴含的因果关系(非时间序
列因果关系),从数据中学习得到的也均是一般变量之间的因果关系.针对这些情况,结合时间序列预处理、时间序列
变量排序、转换数据集构建和局部贪婪打分-搜索等进行时间序列的因果关系学习;再将包括分段在内的时间序列
预处理、时间序列段的因果关系结构学习、因果关系结构数据集构建、因果关系变量排序和局部贪婪打分-搜索
等相结合,来进行元因果关系(因果关系变量之间的因果关系)学习,从而实现两个层次的时间序列因果关系学习,为
进一步的量化因果分析奠定了基础.分别使用模拟、UCI 和金融时间序列数据进行实验与分析,实验结果显示,基于
贝叶斯网络能够有效地进行时间序列的因果关系和元因果关系学习.
关键词: 时间序列;因果关系;贝叶斯网络;转换数据集;结构数据集
中图法分类号: TP181

中文引用格式: 王双成,郑飞,张立.基于贝叶斯网络的时间序列因果关系学习.软件学报,2021,32(10):30683084. http://www.
jos.org.cn/1000-9825/6012.htm
英文引用格式: Wang SC, Zheng F, Zhang L. Learning causal relationship from time series based on Bayesian network. Ruan
Jian Xue Bao/Journal of Software, 2021,32(10):30683084 (in Chinese). http://www.jos.org.cn/1000-9825/6012.htm

Learning Causal Relationship from Time Series Based on Bayesian Network
1
1,2
WANG Shuang-Cheng , ZHENG Fei , ZHANG Li 1
1 (School of Information Management, Shanghai Lixin University of Accounting and Finance, Shanghai 201620, China)
2 (Institute of Data Science and Interdisciplinary Studies, Shanghai Lixin University of Accounting and Finance, Shanghai 201209, China)

Abstract: Bayesian network is a powerful tool for studying the causal relationship between variables. Causal learning, based on
Bayesian network, consists of two parts: structure learning and parameter learning, while structural learning is the core of causal learning.
At present, Bayesian network is mainly used to discover the causality in non-time series data (non-time series causality) and what is
learned from the data is the causal relationship between general variables. In this study, the causality of time series is learned by time
series preconditioning, time series variable sorting, construction of transformation data set, local greedy search-scoring, and so on.
Combining the time series preconditioning including segmentation, the structure learning of causal relationship for time series segments,
the construction of causality structure data set, the variable sorting of causal relationship, local greedy search-scoring, maximum
likelihood parameter estimation, etc., meta causal relationship (used to study the randomness of causal relationship) is established. Thus,
two levels of causality learning can be realized, and the foundation is laid for further quantitative causal analysis. Experiments and
analyses are carried out by using simulation, UCI, and finance time series, the results verify the validity, reliability, and practicability of
learning causal relationship and Meta causality based on Bayesian network.

 基金项目: 国家社会科学基金(18BTJ020)
Foundation item: National Social Science Foundation of China (18BTJ020)
收稿时间: 2018-10-23; 修改时间: 2019-11-06; 采用时间: 2020-01-20; jos 在线出版时间: 2021-01-15