一种基于因果强度的局部因果结构主动学习方法

摘要

Causal structure learning is an important causal knowledge discovery method to disclose the nature of causal interactions in the Bayesian Networks. The causal relations are difficult to be discovered by only using observation data. On the other hand, actually, we are often only interested in local causal structure about a target variable. This paper presented a local causal structure learning method by integrating feature selection into intervention called a local causal structural active learning based on causal power (CSI-LCSL). CSI-LCSL integrated the dividing structure ability of Markov blanket and causal discovery ability of intervention learning. Firstly, under the faithfulness assumption, CSI-LCSL utilized HITON-MB algorithm to obtain the Markov blanket of interested variable for generating a local model Then, we selected a intervention variable from the local model by using norrsys entropy to generate interventional data by perfect experiments. Finally, we used an exact method algorithm to obtain a local causal structure of the interested variable by combining observational data and interventional data. A series of comparative experiments on two standard Bayesian networks show that our method has excellent learning accuracy.%因果结构学习是贝叶斯网络学习中一种重要的结构学习方法,因果关系揭示了系统要素作用的本质.由于仅利用观测数据很难准确地发现变量间的因果关系,且通常人们仅关心网络中关于某一变量的局部因果关系,因此针对难以从观测数据中仅获取所感兴趣的变量的局部因果结构的问题,提出了一种局部结构学习方法,即一种基于因果强度的局部因果结构主动学习方法(CSI-LCSL).CSI-LCSL方法融合了马尔可夫毯的结构划分能力和扰动学习的因果发现能力,并且引入了因果强度进行扰动结点的选择.利用HITON_MB算法寻找目标结点的马尔可夫毯,生成关于目标结点的局部模型；然后,利用不对称信息熵对局部模型中的每一结点进行因果强度分析,选取因果强度值较大的结点进行扰动,生成扰动数据；进而,联合扰动数据和观测数据利用准确方法(exact method)学习边的后验概率,从而获得一个关于目标结点的局部因果网络.利用结构信息熵对CSI-LCSL方法的学习结果进行评估.在标准网络上的实验结果证实了CSI-LCSL算法的有效性.