基于信息论的贝叶斯网络结构学习算法研究

发布时间：2018-07-11 17:39

本文选题：贝叶斯网络 + 信息论　；参考：《东华大学》2017年硕士论文

【摘要】：贝叶斯网络模型是不确定性知识表示和建模领域中最经典、最有效的模型之一。贝叶斯网络自然地将概率论与图论相结合,既能反映出观测数据的内在规律,其图形化的表示方式又具有形象直观的特点,且与神经网络相比为白盒模型,具有良好的可理解性和逻辑性,因此在机器学习、人工智能、生物医疗、故障分析等方面得到了广泛的应用。本文首先对贝叶斯网络模型进行概述,介绍了贝叶斯网络的研究背景及意义、研究历史及现状以及贝叶斯网络的研究前沿。然后按照逻辑的递进顺序详细阐述了信息论的基本知识和贝叶斯网络模型的相关理论,对于贝叶斯网络的学习方法进行了详细介绍,包括两种经典的参数学习方法和两种经典的结构学习方法,并对与贝叶斯网络联系非常紧密的因果关系理论进行了简单介绍。针对连续型随机变量的贝叶斯网络结构学习问题,本文提出了基于条件熵和高斯过程网络的贝叶斯网络结构局部打分法CE-GP,将互信息最大化原则转化为条件熵最小化原则,使用条件熵和高斯过程网络在贝叶斯网络结构学习的过程中为每个节点确定双亲节点集合时进行打分。通过在模拟数据集和酿酒酵母基因数据集上的实验表明,CE-GP打分法能够避免将连续型变量离散化带来的信息丢失,计算简单,准确度高,与本文提出的一种基于阈值的改进K2启发式算法相结合,能够以较小代价获取拟合较优的局部贝叶斯网络结构。针对大规模数据集的贝叶斯网络结构学习问题,本文提出了基于信息论、禁忌搜索和赤池信息准则的贝叶斯网络结构学习算法ITA。首先使用基于信息论的贝叶斯网络降维算法对非目标变量进行筛选,获取与目标变量密切相关的因素,作为贝叶斯网络结构中的节点,然后使用以赤池信息准则作为评分函数,禁忌搜索作为启发式方法的学习算法建立贝叶斯网络模型,以获得全局最优解。通过在SPECT和SIDO数据集上的实验表明,ITA算法能够准确获取领域中的核心因果关系,得到简洁直观的贝叶斯网络,同时具有较低的时间复杂度,是生物医学领域进行大规模不确定性知识表达的一种有效解决方案。
[Abstract]:The Bias network model is one of the most classic and most effective models in the domain of uncertain knowledge representation and modeling. The Bias network naturally combines the probability theory with the graph theory, which can not only reflect the inherent laws of the observation data, but also have the visual and visual characteristics of the graphical representation and the white box model compared with the neural network. It has good understandability and logic, so it has been widely used in machine learning, artificial intelligence, biological medical treatment, fault analysis and so on. This paper first summarizes the Bayesian network model, introduces the research background and significance of Bayesian network, the history and status of the research, and the research frontier of Bayesian Networks. The basic knowledge of information theory and the related theory of Bias network model are elaborated in detail according to the progressive sequence of logic, and the learning methods of Bias network are introduced in detail, including two classical parameter learning methods and two classical structural learning methods, and a very close causal relationship with Bayesian Juliu network. In view of the Bayesian network structure learning problem of continuous random variables, this paper proposes a local scoring method CE-GP based on conditional entropy and Gauss process network, which transforms the principle of mutual information maximization into the principle of conditional entropy minimization, and makes use of conditional entropy and Gauss process network in Bayesian networks. Through the experiments on the simulated dataset and the Saccharomyces cerevisiae data set, the CE-GP scoring method can avoid the loss of information caused by the discretization of continuous variables, and the calculation is simple and accurate. Combining the K2 heuristic algorithm, we can obtain better fitting local Bayesian network structure at a small cost. In view of the Bayesian network structure learning problem of large-scale data sets, this paper proposes a Bayesian network structure learning algorithm based on information theory, tabu search and red pool information criteria, ITA. first uses information based shellfish. The Juliu network reduction algorithm filters the non target variables and obtains the factors closely related to the target variables. As the nodes in the Bias network structure, the Bayesian network model is established by using the Chek pool information criterion as the scoring function and the tabu search as a heuristic method to obtain the global optimal solution. The experiments on SPECT and SIDO datasets show that the ITA algorithm can accurately obtain the core causality in the field, and get a concise and intuitive Bayesian network with a lower time complexity. It is an effective solution to the large-scale uncertainty knowledge expression in the biomedical field.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【相似文献】