基于观测数据的时间序列因果推断综述  

Overview of observational data-based time series causal inference

作　　者：曾泽凡 陈思雅 龙洗 金光[1] ZENG Zefan;CHEN Siya;LONG Xi;JIN Guang(School of Systems Engineering,National University of Defense Technology,Changsha 410073,China;School of Aerospace Science,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]国防科技大学系统工程学院,湖南长沙410073 [2]国防科技大学空天科学学院,湖南长沙410073

出　　处：《大数据》2023年第4期139-158,共20页Big Data Research

摘　　要：数据存储量的扩大和计算能力的提升为基于观测数据推断时间序列的因果关系开辟了新途径。在时间序列因果推断的基本性质和研究现状的基础上,系统梳理了5种基于观测数据的时间序列因果推断方法,即Granger因果分析方法、基于信息论的方法、因果网络结构学习算法、基于结构因果模型的方法和基于非线性状态空间模型的方法。然后,根据不同应用场景的数据特点,结合方法的功能和适配性,对基于观测数据的时间序列因果推断方法在经济金融、医疗和生物学、地球系统科学和其他工程领域的典型应用进行了简要介绍。最后,结合时间序列因果推断的重难点问题,比较5种方法的优缺点,分析下一步研究重点,展望未来的研究方向。With the increase of data storage and the improvement of computing power,using observational data to infer time series causality has become a novel approach.Based on the properties and research status of time series causal inference,five observational data-based methods were induced,including Granger causal analysis,information theory-based method,causal network structure learning algorithm,structural causal model-based method and method based on nonlinear state-space model.Then we briefly introduced typical applications in economics and finance,medical science and biology,earth system science and other engineering fields.Further,we compared the advantages and disadvantages and analyzed the ways for improvement of the five methods according to the focus and difficulties of time series causal inference.Finally,we looked into the future research directions.

关 键 词：时间序列 因果推断 GRANGER因果分析 信息熵 贝叶斯网络 结构因果模型 非线性状态空间模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]