【作者】 李建；

【导师】 黄少滨；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2011， 硕士

【摘要】 随着社会信息化的发展,人类现在以及未来都面临着信息爆炸的问题,对于数据的分析和处理正变得日益困难。在此背景下,聚类分析技术应运而起,并得到了蓬勃发展,很多聚类算法被相继提出。但是,任何聚类算法都是建立在一定的假设基础之上的,由于不可能所有的数据集都满足同一假设,因此在许多情况下单一聚类算法无法取得满意的结果。聚类融合通过把具有一定差异性的聚类成员进行组合,能够得到比单一算法更为优越的结果,并兼有鲁棒性、可并行性等优点,所以迅速得到了国内外学者的重视,融合学习也被人称为机器学习领域未来四个重要的研究方向之一。本论文研究的主要内容就是聚类融合。首先,结合“最近邻”的思想,本论文提出了一种基于自适应最近邻的聚类融合算法—NNCE。该算法能够根据数据分布密度的不同,为每一个数据点自动选择合适的最近邻选取范围,较好地解决了基于KNN的算法中存在的数据点的最优最近邻数量K需要实验确定的问题并进一步提高了聚类效果。其次,本论文扩展了“核心群”的思想,给出了“绝对核心群”和“相对核心群”的概念,并利用ANN思想产生一个簇的核心群,给出了基于核心群的K-means优化算法—RCBK-means,该算法能够比较好地解决经典K-means算法因随机选取初始中心点而导致的聚类结果的局部最优化问题。最后,本论文还给出对一种基于质心的聚类融合算法CBEC的改进方法--RCC-CBEC。该方法通过使用簇中相对核心群的质心来代替簇的质心,能够更精确地代表这个簇,从而得到更好的全局质心,提高最终的聚类质量。更多还原

【Abstract】 With the development of human society, we will face the problem of information explosion in the future. In this context, the technology of cluster analysis has been vigorous development.People have proposed a lot of clustering algorithms. However, any clustering algorithm has a assumption about data distribution, resulting in that a single clustering algorithm can not achieve satisfactory result in many cases. Clustering ensemble algorithm can be more advantageous than a single algorithm.Clustering ensemble algorithms have been successfully applied in many areas and are considered as one of the four important future research directions for machine learning.The main content of this thesis is about the clustering ensemble. First, using the concept of nearest neighbor, this thesis proposed a new clustering ensemble algorithm-ANNCE.The advantages of this algorithm are:First, it combin the results of multiple clustering algorithms and has better stability and accuracy; Secondly, according to the different density of data, our algorithms can effectively automatically select a different number of nearest neighbors for each data point and get a good result. IN the second part of this thesis, we propose two concepts-"absolute core cluster"(ACC) and "relative core cluster"(RCC), and we proposed RCBK-means (Relative Core Cluster Based K-means) algorithm to solve the local optimization problem. In addition, we also used RCC to improve result of the CBEC algorithm and.proposed RCC-CBEC algorithm.更多还原