基于MFO的贝叶斯网络结构学习及应用

发布时间：2018-04-26 00:39

本文选题：贝叶斯网络 + 结构学习　；参考：《中国科学技术大学》2017年硕士论文

【摘要】：信息时代的蓬勃发展使人们积累了大量数据,将贝叶斯网络用于数据挖掘,从海量数据中挖掘蕴含的知识、逻辑,抽取具有使用价值的信息,具有重要意义。贝叶斯网络是图论和概率论相结合的图形化网络模型,该模型直观明了,在不完备数据、不确定信息上具有较强的处理能力,广泛应用于数据分析以及不确定性信息处理等领域,值得研究推广。构建贝叶斯网络涉及结构和参数的学习,其中结构学习是技术关键,直接关系到参数学习的结果,继而影响应用效果,研究结构学习算法具有很强的必要性。本文主要工作如下:1.针对主流结构学习方法——基于评分搜索的方法普遍存在精确度不高、结构返回不稳定、容易陷入局部最优等问题,本文首次将飞蛾-烛火优化(Moth-Flame Optimization,MFO)算法引入结构学习,提出了基于MFO的结构学习算法(Bayesian Network Structure Learning using MFO,BN-MFO)。BN-MFO 保留了MFO的整体框架,通过借鉴遗传算法的杂交、变异等操作,替换了 MFO中的曲线位置更新方法。在变异操作时,参考节点间互信息,对不同的节点采用不同的变异动作,使搜索趋向于数据蕴含的潜在结构。实验研究了 BN-MFO中评分函数和搜索策略的关系及R函数关系对于收敛性能的影响,分析了 BN-MFO的有效性。在经典的Cancer网络和Asia网络上的对比实验结果表明BN-MFO普遍优于同类型的对比算法,具有较强的优越性。2.将贝叶斯网络应用于银行营销数据分析中,运用BN-MFO学习网络结构,在实际应用中检验了算法有效性。贝叶斯网络具有分类能力,本文实验对比了和KNN、SVM的分类准确率,效果较优,从侧面反映了 BN-MFO的有效性。为了便于模型的使用,还设计了基于Matlab的GU1软件。综上所述,本文研究解决了基于评分搜索的结构学习方法中普遍存在的问题,给出了一种稳定返回最优结构集合的方法。本文的研究工作,拓宽了贝叶斯网络模型的构建方式,对推动贝叶斯网络理论发展和拓宽贝叶斯网络的应用领域有一定意义。
[Abstract]:The vigorous development of the information age has made people accumulate a lot of data. It is of great significance to use the Bias network for data mining, mining the knowledge, logic, and extracting the useful information from the massive data. The Bias network is a graphical network model combining the graph theory with the probability theory. The model is intuitive and incomplete. Data and uncertain information have strong processing ability, which are widely used in data analysis and uncertainty information processing. It is worth studying and popularizing. The construction of Bayesian networks involves the learning of structure and parameters. Structure learning is a key technology, which is directly related to the results of reference learning, and then influences the application effect and research structure. The main work of this paper is as follows: 1. in view of the mainstream structure learning method, the method based on the score search is generally not high in accuracy, the structure is unstable and easy to fall into the local optimal problem. In this paper, the Moth-Flame Optimization (MFO) algorithm is introduced to structural learning for the first time. The MFO based architecture learning algorithm (Bayesian Network Structure Learning using MFO, BN-MFO).BN-MFO retained the overall framework of MFO, replacing the curve position updating method in MFO by using the hybrid and mutation operations of genetic algorithms. In the mutation operation, the information between the reference nodes and the different nodes are different. The variation action that makes the search tend to the potential structure of the data implication. The experiment studies the relationship between the scoring function and the search strategy in BN-MFO and the effect of the R function relationship on the convergence performance, and analyzes the validity of the BN-MFO. The comparison experiment results on the classical Cancer network and the Asia network show that BN-MFO is generally superior to the same type. The algorithm has strong superiority.2. to apply the Bias network to the analysis of bank marketing data. Using BN-MFO to learn network structure, the effectiveness of the algorithm is tested in the practical application. The Bias network has the ability of classification. This paper compares the experiment with KNN, and the accuracy of the classification of SVM is better, and the effectiveness of BN-MFO is reflected from the side. In order to facilitate the use of the model, the GU1 software based on Matlab is also designed. To sum up, this paper studies and solves the common problems in the structure learning method based on the score search, and gives a method to return the optimal set of optimal structures. The development of Juliu network theory and broadening the application area of Bayesian network are of some significance.

【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【参考文献】