基于非时序观察数据的因果关系发现综述

发布时间：2019-08-15 07:16

【摘要】：探索和发现事物间的因果关系是数据科学的一个核心问题,其中蕴含着丰富的科学发现机会和巨大的商业价值.基于非时序观察数据的因果关系发现方法能够从被动观察获得的数据中发现变量之间的因果关系,因而在各领域有广泛应用.这一类方法在过去三十年取得很大进展,已经成为因果关系发现的重要途径.文中从因果关系方向推断、高维数据上的误发现率控制和不完全观察数据上的隐变量检测这三个研究热点出发,对现有的因果关系模型与假设、基于约束的方法、基于因果函数模型的方法和混合型方法这三大类方法,验证与测评涉及的数据集及工具等方面进行了详尽的介绍与分析.基于约束的方法主要包括因果骨架学习和因果方向推断两个阶段:首先基于因果马尔可夫假设,采用条件独立性检验学习变量之间的因果骨架,然后基于奥卡姆剃刀准则利用V-结构确定因果方向,典型的算法有Peter-Clark算法、Inductive Causation等,这类方法的主要不足是存在部分无法判断的因果关系方向,即存在Markov等价类难题.基于因果函数模型的方法则基于数据的因果产生机制假设,在构建变量之间的因果函数模型的基础之上,基于噪声的非高斯性、原因变量与噪声的独立性、原因变量分布与因果函数梯度的独立性等因果假设推断变量之间的因果关系方向,典型的算法有针对线性非高斯无环数据的Linear NonGaussian Acyclic Model算法、针对后非线性数据的Post-NonLinear算法、适用于非线性或离散数据的Additive Noise Model等,这类方法的主要不足是需要较为严格的数据因果机制假设,且Additive Noise Model等方法主要适用于低维数据场景.混合型方法则希望充分发挥基于约束的方法和基于因果函数类方法的优势,分别采用基于约束的方法进行全局结构学习和基于因果函数模型进行局部结构学习和方向推断,典型的算法有SADA、MCDSL等,理论分析较为不足是这类方法目前遇到的主要困难.最后,文中还基于研究现状分析讨论了因果方向推断、高维数据上的误发现率控制、隐变量发现、与机器学习的关系等未来可能的研究方向.
[Abstract]:Exploring and discovering the causality between things is a core problem of data science, which contains rich opportunities for scientific discovery and great commercial value. The causality discovery method based on non-sequential observation data can find the causality between variables from the data obtained by passive observation, so it has been widely used in various fields. This kind of method has made great progress in the past three decades and has become an important way to discover causality. In this paper, it is inferred from the direction of causality, the false detection rate control on high-dimensional data and the detection of hidden variables on incomplete observation data are introduced and analyzed in detail, such as the existing causality model and hypothesis, the method based on constraint, the method based on causality function model and the mixed method, and the data sets and tools involved in verification and evaluation. The method based on constraint mainly includes two stages: based on causality Markov hypothesis, the causality skeleton between learning variables is tested by conditional independence, and then the causality direction is determined by V-structure based on Occam razor criterion. The typical algorithm is Peter-Clark algorithm, Inductive Causation. The main deficiency of this method is that there are some causality directions that can not be judged. That is, there is a problem of Markov equivalence. The method based on causality function model is based on the hypothesis of causality generation mechanism of data. On the basis of constructing causality function model between variables, the causality direction between variables is inferred based on the non-Gaussian property of noise, the independence of cause variable and noise, the distribution of cause variable and the independence of causality function gradient. The typical algorithm has Linear NonGaussian Acyclic Model algorithm for linear non-Gaussian acyclic data. The Post-NonLinear algorithm for post-nonlinear data is suitable for Additive Noise Model of nonlinear or discrete data. The main deficiency of this method is that it needs more strict hypothesis of data causality mechanism, and Additive Noise Model and other methods are mainly suitable for low-dimensional data scenarios. The hybrid method hopes to give full play to the advantages of constraint-based method and causality function class method, respectively, using constraint-based method for global structure learning and local structure learning and direction inference based on causality function model. The typical algorithms are SADA,MCDSL and so on. The shortage of theoretical analysis is the main difficulty encountered by this kind of methods at present. Finally, based on the analysis of research status, this paper also discusses the possible future research directions, such as causality inference, false detection rate control on high-dimensional data, hidden variable discovery, and the relationship with machine learning.
【作者单位】：广东工业大学计算机学院;卡内基梅隆大学哲学系;佛山科学技术学院数学与大数据学院;
【基金】：NSFC-广东联合基金(U1501254) 国家自然科学基金(61572143) 广东省杰出青年科学基金(2014A030306004)资助~~
【分类号】：TP18

【参考文献】