【作者】 陈阳；

【导师】 罗可；

【作者基本信息】 长沙理工大学 ， 计算机技术（专业学位）， 2014， 硕士

【摘要】 在这个“信息爆炸”的时代,人们经常面对着海量的数据,例如海量文本数据、Web数据、多媒体数据等,而这些数据中隐藏着大量的对人们有用的重要信息,如何用某种有效手段,从这些海量数据中提取出这些潜在有用的、可以用于提供决策支持的信息,数据挖掘便应运而生了。聚类作为数据挖掘的一个重要分支,对人们的日常生活都产生非常重要的影响。现在聚类主要的技术可以分为：划分方法、层次方法、基于密度方法、基于网格方法和基于模型方法。本文的主要研究工作包括下面两个方面；1.从基于划分的K-medoids聚类算法的理论基础入手,探讨K-medoids算法的不足,提出了一种基于改进禁忌搜索的K-medoids聚类算法。禁忌搜索算法是模拟人类短暂记忆的一种优化算法,具有很强的全局搜索能力。针对禁忌搜索算法比较依赖初始解的缺点,将粒计算和最大距离积算法相结合,将改进后的禁忌搜索算法提高了K-medoids聚类的准确率,有较强的稳定性。通过仿真实验,验证了该方法的可行性。2.根据高维空间数据的特性,传统的低维的距离计算相似性的方法不合适用在高维空间中,提出了一种新的相似性度量方法。通过计算对象之间的相似度得到相似矩阵,结合层次聚类的思想,自底向上的对数据进行聚类分析。通过有效性分析,论证了该方法可行性。更多还原

【Abstract】 In the era of "information explosion", we are often faced with the massive data, such as the massive text data, Web data, multimedia data, and those data hide a lot of useful and important information. How to extract that potential information which can be used to support decision using some effective methods leads to the birth of data mining. As one of the important branch of data mining, clustering has a great impact on people’s daily life. The main technologies of existing clustering can be divided into these category:Partitioning method, hierarchical method, density-based method, grid-based method and model-based method.The main research work of this thesis includes the following two aspects:1. Starting from the theoretical basis of partition based K-medoids clustering algorithm and discussing the deficiency of K-medoids algorithm, we proposed an K-medoids clustering algorithm based on improved Tabu Search. Tabu search algorithm is an optimization algorithm simulating human short-term memory making it has a strong global searching ability. View of the shortcomings that tabu search comparatively relies on the initial solution, we combined the granular computing with maximum distance product method to improve tabu search, which improved the precision of K-medoids clustering with a strong stability. Through the experiment, we verified the feasibility of the method.2. According to the characteristics of high dimensional spatial data, and the distance similarity computing method in low dimensional is not suitable in the high dimension space, we proposed a new similarity measure method. Having gotten the similar matrix through computing the similarity between objects, we undertook clustering analysis inspired by hierarchical clustering bottom-up notion. Through the analysis on validity, we proved the feasibility of this method.更多还原