贝叶斯网络的因果隐变量发现及其应用研究

发布时间：2018-03-22 22:06

本文选题：贝叶斯网络　切入点：隐变量　出处：《合肥工业大学》2014年硕士论文　论文类型：学位论文

【摘要】：隐变量指的是观察不到的变量,包含了关于事物本质的关键信息,这些变量能够简化结构,汇聚显变量之间的依赖关系。隐变量的发现有利于人们对于事物真实状态和特性的认知,贝叶斯网络的隐变量学习是数据挖掘和知识发现研究领域中的一个重要研究内容。论文结合因果分析技术和不确定性技术研究隐变量发现问题,并基于复杂系统从能量的角度开展隐变量的应用研究,项目的研究工作具有较高的现实意义及应用价值。具体研究工作如下：第一,结构分析的隐变量发现方法难以有效地发现隐变量且可解释性较差,提出一种基于局部因果关系分析的隐变量发现算法(LCAHD)。LCAHD算法的基本思想：首先寻找目标变量的马尔科夫毯提取局部依赖结构,然后基于扰动学习获得扰动数据,联合扰动数据和观测数据学习局部依赖结构中的因果关系；进而,利用给出的因果结构熵计算模型对局部因果结构中因果关系的不确定性进行度量,并利用隐变量和因果关系不确定性之间的相关性判定条件,确定隐变量的存在性；最后,利用因果非对称信息熵对隐变量的重要性进行衡量,并给出隐变量发现算法。第二,股市中的数据具有海量、多源等特点,使得数据的维数较高,导致预测的准确性下降,针对这一缺点,提出了基于特征融合的隐变量学习及在金融网络的研究(LHFF)。该算法的基本思想是：首先,收集影响股市能量的特征,利用互信息对特征之间的关联度进行计算,并依据关联度的大小进行特征提取；然后,利用对提取的特征赋权值,进行特征融合,融合后成为隐变量即股市的能量,建立最终的能量计算模型；最后,利用能量计算模型计算能量的大小,根据能量的大小对大盘指数进行预测分析。根据股市的实践案例表明该算法具有很强的实用性。在标准网络和股票网络进行了算法的实验,结果表明该方法能准确地确定隐变量的位置,且具有较好的解释性。在金融网络上的实验效果表明,隐变量在实际生活领域中的广泛应用性。
[Abstract]:Implicit variables are unobserved variables that contain key information about the nature of things that simplify the structure. The discovery of hidden variables is beneficial to people's cognition of the real state and characteristics of things. The hidden variable learning of Bayesian networks is an important research content in the field of data mining and knowledge discovery. And based on the complex system from the perspective of energy to carry out the application of hidden variables, the research of the project has a higher practical significance and application value. The specific research work is as follows: first, It is difficult to find hidden variables effectively in structural analysis and can not be explained effectively. A new implicit variable discovery algorithm based on local causality analysis (LCAHDN. LCAHD) is proposed. Firstly, the Markov blanket of the target variable is found to extract the local dependency structure, and then the disturbance data is obtained based on the perturbation learning. The joint disturbance data and observation data are used to study the causality in the local dependency structure, and then, the uncertainty of the local causality in the local causality structure is measured by using the given entropy calculation model of the causality structure. The existence of hidden variables is determined by using the correlation judgment condition between hidden variables and causal uncertainties. Finally, the importance of hidden variables is measured by causal asymmetric information entropy, and the algorithm of hidden variable discovery is given. The data in the stock market have the characteristics of mass, multi-source and so on, which make the dimension of the data higher and lead to the decline of the accuracy of prediction. The hidden variable learning based on feature fusion and its research in financial network are proposed. The basic idea of this algorithm is as follows: firstly, the features affecting the energy of the stock market are collected, and the correlation degree between the features is calculated by mutual information. And according to the magnitude of the correlation degree of feature extraction; then, using the extracted feature weighting value, the feature fusion, fusion into a hidden variable, that is, the energy of the stock market, establish the final energy calculation model; finally, The energy calculation model is used to calculate the energy, and the large market index is forecasted and analyzed according to the energy size. The practical cases of the stock market show that the algorithm is very practical. The experiments of the algorithm are carried out in the standard network and the stock network. The results show that this method can accurately determine the location of hidden variables and has a good explanation. The experimental results on financial networks show that the hidden variables are widely used in real life.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：F832.51;F224

【相似文献】