进化神经网络聚类技术及其在数据挖掘中的应用

【作者】 王石；

【导师】 李玉忱；

【作者基本信息】 山东大学 ， 计算机软件与理论， 2005， 硕士

【摘要】 数据挖掘技术是解决数据丰富而知识贫乏的有效途径,当属信息科学领域的前沿研究课题之一,有关的研究和应用极大提高了决策支持的能力,已被公认为是数据库研究中一个极富应用前景的领域。聚类分析是数据挖掘的一个重要内容,包括基于划分、层次、密度、网格和模型的分析思想。本文讨论了聚类技术当前的研究现状,研究和比较了各种聚类算法的各自特点及不足之处。特别地,对于基于模型思想的SOM神经网络聚类技术做了详细的探讨。 SOM神经网络聚类是典型的基于模型思想的聚类方法,是脑认知科学和数据挖掘的有机结合,与实际的大脑处理知识有很强的理论联系。但该方法在效率、结果最优性和网络模型结构参数难以确定方面均存在不足之处。 基于遗传算法的进化神经网络是近几年发展起来的一个跨学科的新的非常活跃的研究领域,它包含三个层次的进化策略,即连接权的进化、网络结构的进化和学习规则的进化。进化算法思想为神经网络系统的设计、系统性能的改进、神经网络的集成等提供了新的研究思路。 在传统SOM神经网络聚类的基础上,结合进化神经网络理论,本文提出一种新的基于进化SOM神经网络的聚类算法ESOM。其思想是利用基于遗传算法的进化理论对SOM网络的权值向量集进行优化,同时采用神经元分裂的思想对网络模型结构进行递归扩展优化,从而得到最优网络结构。 本文详细描述了该算法的实现过程,讨论了聚类精度的度量标准,并利用其来控制网络的生长和评估聚类结果。本文给出了算法的正确性和时间复杂度分析,并通过仿真试验比较了ESOM算法和经典SOM聚类方法以及基于遗传算法的SOM聚类方法的性能和聚类精度,结果说明ESOM克服了传统SOM神经网络的上述不足之处,而相对于基于遗传算法的SOM聚类方法能得到更好的聚类结果和执行效率,这主要是由于对其网络结构的优化造成的结果。 最后,本文给出了一个决策支持系统的系统框架,并对算法在该系统中的应用做了简单介绍。更多还原

【Abstract】 Data mining technology is an effective approach to resolve the problem of abundant data and scanty information. It currently is the research frontier within the information science field. The related researches and applications have greatly improved the ability for decision supporting. It has been deemed to a field that has broad prospect of application in database research. As an important part of DM, Clustering Analysis consists Partitioning Method, Density-Based Method, Grid-Based Method, and Model-Based Method. This paper discusses the developing status of Clustering technology recently, compares the characteristics of all these arithmetic and points out the advantages and disadvantages of them. Specially, this paper makes deeper research on SOM-Based Clustering Technology.SOM-Based Clustering Technology is an integration of neural technology and DM. It has a strong connection with the process of brain. But this method has many shortcomings in respect of efficiency, optimization and the parameter, which determine the structure of the network.Evolutionary Neural Network is an active research field recently these days. Its idea of evolutionary includes three levels: the evolutionary of connection weights, the evolutionary of net structure and the evolutionary of study rules. Evolutionary Neural Network provides a new idea for the design of Artificial Neural Network, the improvement of performance of Neural System and the integration of Networks.Based on the traditional SOM-based Clustering technology and enlighten by the Evolutionary Theory, this paper introduces a new clustering methods based on Evolutionary SOM. It evolves the weights of the cells and optimizes the structure of the network by "splitting" the cells.This paper describes the arithmetic in detail, discusses the criterion of quality of Clustering and controls the growth of the network using it. The experimentationshows it overcomes the shortcomings of SOM Method listed above. And because of optimizing of the network structure, ESOM is more efficiency and accurate than genetic-based Method.Lastly, this paper introduces a system framework of a Decision Support System, and discusses the operation of the method in this system.更多还原