贝叶斯网络局部因果结构学习方法研究及其应用

发布时间：2019-01-03 17:16

【摘要】：在过去的数年间，，大量的研究集中在基于数据驱动的方式发现因果贝叶斯网络。因果关系揭示了系统要素作用的本质，因果结构学习已成为贝叶斯网络学习中一种重要的结构学习方法。但是从数据中学习全局贝叶斯网络是一个NP难题，并且人们通常仅关心其感兴趣的变量的局部因果结构。因此，学习贝叶斯网络中关于目标变量的局部因果结构成为该领域的一个重要研究内容。局部因果结构学习方法可以应用于发现经济领域的因果知识。发现目标变量的局部因果结构是一个充满挑战性的问题，同时也具有重要的科学意义和较高的应用价值。论文中，将针对目标变量局部因果结构学习展开一系列的探讨。局部因果结构学习主要包括两个方面：一是，从数据中获取关于目标变量的局部模型；二是，发现局部模型中变量间的因果关系。针对这两个方面的研究内容，本文所开展的研究工作具体如下：第一，为了能够有效的学习网络中目标变量的局部因果结构，提出了基于扰动学习的局部因果结构学习方法（I-LCSL方法）。首先，利用马尔可夫毯的结构划分能力从观测数据中获取目标变量的局部模型。其次，利用扰动学习的因果发现能力，从局部模型中随机选择一个扰动变量进行扰动，进而获取一组扰动数据。最后，联合观测数据和扰动数据，利用动态规划算法学习关于变量的局部因果结构。利用边的错误率对I-LCSL方法学习到的局部因果结构进行评估。第二，针对现有方法都难以确定扰动节点的缺点，提出一种基于因果强度的局部因果结构主动学习方法（CSI-LCSL方法）。首先，寻找目标节点的马尔可夫毯生成关于目标节点的局部模型；其次，利用不对称信息熵对局部模型中的每一节点进行因果强度分析，选取因果强度值较大的节点进行扰动，生成扰动数据；再次，联合扰动数据和观测数据学习边的后验概率，从而获得一个关于目标节点的局部因果网络。最后，利用结构信息熵对CSI-LCSL方法的学习结果进行评估。通过现实生活中的数据，对本文所提出的方法进行了综合评估。实验表明，本文提出的方法能够有效发现目标变量的局部因果结构，且学习结果优于当前的其它算法。最后，我们使用本文提出的局部因果结构学习方法在金融行业板块数据集上发现房地产行业和煤炭石油行业的局部因果结构。
[Abstract]:In the past few years, a large number of studies have focused on the discovery of causal Bayesian networks in a data-driven way. Causality reveals the essence of the function of system elements. Causal structure learning has become an important structural learning method in Bayesian network learning. But learning global Bayesian networks from data is a NP problem, and people usually only care about the local causal structure of the variables of interest. Therefore, learning the local causal structure of target variables in Bayesian networks is an important research topic in this field. The local causality structure learning method can be applied to discover the causality knowledge in the economic field. Finding out the local causal structure of the target variable is a challenging problem and has important scientific significance and high application value. In this paper, a series of studies will be carried out on the local causal structure learning of target variables. The learning of local causality structure mainly includes two aspects: one is to obtain the local model about the target variable from the data, the other is to find the causality between the variables in the local model. For these two aspects of the research work, the research work carried out in this paper is as follows: first, in order to effectively learn the local causal structure of the target variables in the network, A local causal structure learning method based on disturbance learning (I-LCSL method) is proposed. Firstly, the local model of the target variables is obtained from the observed data by using the structural partition ability of Markov blanket. Secondly, using the causality discovery ability of disturbance learning, a disturbance variable is randomly selected from the local model for disturbance, and then a set of disturbance data is obtained. Finally, the dynamic programming algorithm is used to study the local causal structure of variables. The error rate of edges is used to evaluate the local causality structure learned by I-LCSL method. Secondly, an active learning method based on causality strength for local causal structure (CSI-LCSL method) is proposed to overcome the disadvantages of the existing methods which are difficult to determine the disturbance nodes. Firstly, the local model of the target node is generated by finding the Markov blanket of the target node. Secondly, the asymmetric information entropy is used to analyze the causality strength of each node in the local model. Thirdly, a local causal network for the target node is obtained by combining the posteriori probability of the learning edge between the perturbed data and the observed data. Finally, the structural information entropy is used to evaluate the learning results of CSI-LCSL method. Based on the real-life data, the method proposed in this paper is evaluated synthetically. Experimental results show that the proposed method can effectively find the local causal structure of the target variables and the learning results are better than those of other current algorithms. Finally, we use the local causal structure learning method proposed in this paper to find the local causal structure of real estate industry and coal oil industry on the financial sector plate data set.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP181

【相似文献】