高维因果网与高校资产管理的模糊推理研究

发布时间：2018-01-03 21:49

本文关键词：高维因果网与高校资产管理的模糊推理研究　出处：《华南理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：贝叶斯网络作为处理变量间因果关系表达和推理的有效工具,在人工智能与数据挖掘领域受到广泛的应用。但传统的贝叶斯网络也存在以下三个问题,制约进一步的发展。首先是传统贝叶斯网络只能处理离散变量,处理其它类型变量时需要先进行离散化,这样容易产生边缘锐化‖问题。特别是面对模糊变量时,会造成大量的信息丢失,影响推理精度。然后是贝叶斯网络的结构学习是NP-HARD问题,在高维数据下,其网络结构的搜索空间会呈指数增长,常用结构学习算法的效率都会变得低下。最后是传统的结构学习算法无法识别马尔科夫等价类,当搜索空间存在大量的马尔科夫等价类,搜索效率就会很低,且很难收敛到最优解中。针对问题一,本文根据模糊理论对传统贝叶斯网络进行扩展,给出了能兼容模糊变量的混合贝叶斯网络的完整方案。对于问题二,则提出了一种新型的约简组合方案通过把建网问题分割为多子网的构建,降低高维数据的影响。同时该方法在聚类的过程同时确定子网间的连接点,避免了二次搜索,与同类算法相比降低了计算复杂度。聚类的过程使用了基于因果的相似度,降低分网建立对结构质量的负面影响。最后针对问题3,则把基于信息几何的因果推断与爬山法结合,提出了改进的爬山算法IGCI-HC,解决传统结构学习算法无法识别马尔可夫等价类的不足。高校资产管理直接影响高校的建设与发展。影响高校资产管理的因素有多个,具有高维且涉及因素类型多的特点,如何挖掘分析因素间以及资产管理效率间的因果关系并给予知识推理决策辅助,是本文的应用要点。本文根据72所高校的资产管理相关数据,建立了结合约简组合算法的混合因果网,对该问题进行了知识推理分析。实验结果表明,混合贝叶斯网络虽然因模糊变量的处理增加了一些计算复杂度,但是通过结合约简算法与IGCI—HC算法,降低了计算复杂度。另外,由于其采用了模糊概率表示变量的不确定性,解决了边缘锐化的问题。与传统的贝叶斯网络相比,混合贝叶斯网络在网络质量与推理精度上都要更为出色。
[Abstract]:As an effective tool to deal with causality expression and reasoning among variables, Bayesian network is widely used in artificial intelligence and data mining, but the traditional Bayesian network also has the following three problems. First of all, traditional Bayesian networks can only deal with discrete variables, and other types of variables need to be discretized first. Especially in the face of fuzzy variables, a large amount of information will be lost, which will affect the reasoning accuracy. Then the structural learning of Bayesian networks is the NP-HARD problem. In high-dimensional data, the search space of network structure will increase exponentially, and the efficiency of common structural learning algorithms will become low. Finally, the traditional structural learning algorithm can not recognize Markov equivalent class. When there are a large number of Markov equivalence classes in search space, the search efficiency will be very low, and it is difficult to converge to the optimal solution. For the first problem, this paper extends the traditional Bayesian network according to fuzzy theory. A complete scheme of hybrid Bayesian networks which can be compatible with fuzzy variables is presented. For problem two, a new reduction combination scheme is proposed by dividing the problem into multiple subnets. The influence of high-dimensional data is reduced. At the same time, in the process of clustering, the join points between subnets are determined, and the secondary search is avoided. Compared with the similar algorithm, the algorithm reduces the computational complexity. The clustering process uses the similarity based on causality to reduce the negative impact on the structure quality caused by the establishment of the subnet. Finally, the problem 3 is addressed. Then combining the causal inference based on information geometry with the mountain climbing method, an improved mountain climbing algorithm IGCI-HC is proposed. To solve the traditional structural learning algorithm can not identify the shortcomings of Markov equivalents. College asset management directly affects the construction and development of colleges and universities. There are many factors that affect the asset management of colleges and universities. It has the characteristics of high dimension and many types of factors involved. How to mine the causal relationship between factors and asset management efficiency and give knowledge reasoning decision assistance. According to the related data of asset management in 72 colleges and universities, a hybrid causality net combining reduction combination algorithm is established, and the knowledge reasoning analysis of this problem is carried out. The experimental results show that. Although the hybrid Bayesian network increases some computational complexity because of the fuzzy variable processing, it reduces the computational complexity by combining the reduction algorithm with the IGCI-HC algorithm. Because of the uncertainty of fuzzy probability representation variables, the problem of edge sharpening is solved. Compared with traditional Bayesian networks, hybrid Bayesian networks are better in network quality and reasoning accuracy.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP18;G647.5

【相似文献】