演化聚类算法研究及其应用

发布时间：2018-05-17 05:30

本文选题：聚类 + 演化数据　；参考：《扬州大学》2017年硕士论文

【摘要】：聚类作为数据挖掘领域中一种非常有效的数据分析方法,得到了很多学者的研究,在模式识别、图像处理、数据压缩等领域得到了广泛的应用。所谓聚类就是将数据对象分组成为多个类或簇(Cluster),其划分原则是将具有较高的相似度的数据对象间划分到同一个簇中,而相似度误差较大的数据对象应划分到不同的簇中。传统的聚类算法只能针对一些静态数据有很好的处理效果,而对于近年来引起机器学习和数据挖掘领域重点关注的演化数据,更有待学者们的研究。由于演化数据是随时间的推移数据分布会发生变化、有新数据的出现或旧数据的消亡,那么怎样做到使每一时刻上的数据聚类性能尽可能的好,能够基本正确地反映每一时刻的数据分布;通过聚类发掘数据的演化机制,例如聚类的出现、变化、分裂、消失等;还要使得聚类结果在时间上要尽可能平滑,使得当前时刻的聚类结果与前一时刻的聚类结果尽可能的相类似,已有小部分学者进行了研究。本文着重研究演化数据的聚类问题,研究了两种无监督的演化聚类算法和半监督(带约束)的演化聚类算法,并进行了简单的应用。具体研究工作和成果如下:(1)本文提出了基于时间平滑性的演化聚类框架,其框架是在Chakrabarti等人提出的在线式框架基础上进行修改完善得到的。除此之外,本文还对数据间的相似度矩阵作出了公式定义,相似度计算包括两个部分之和:当前时刻数据间的相似度与时间序列上的相似度。最后,并将框架具体应用到标准谱聚类当中,得到两种新的演化谱聚类算法并进行实验验证。(2)本文提出了演化的双层随机游走半监督聚类算法,其算法是针对处理带有约束信息的演化聚类的。原始的静态双层随机游走半监督聚类算法在处理不断变化增加的数据时,会花费大量的时间与内存,并且不能得到很好的效果。本文在双层随机游走半监督聚类算法的基础上很好的利用之前时刻的信息,通过在高层随机游走时求解组件间的两两相似度直接加入前一时刻旧数据信息,大大减少了计算的时间,更好的处理演化半监督数据,能够得到较好的聚类结果。(3)本文设计了一种演化的人脸聚类系统,此系统中的人脸聚类匹配是通过应用本文提出的演化聚类算法来处理的。本系统主要功能包括三个部分:人脸处理(演化聚类)、识别结果显示、文件的管理。
[Abstract]:Clustering, as a very effective data analysis method in the field of data mining, has been studied by many scholars, and has been widely used in the fields of pattern recognition, image processing, data compression and so on. Clustering is the grouping of data objects into multiple classes or clusters. The principle of clustering is to divide data objects with high similarity into the same cluster, while data objects with large similarity errors should be divided into different clusters. The traditional clustering algorithm can only deal with some static data very well, but for the evolutionary data which has attracted the attention of the field of machine learning and data mining in recent years, it needs to be studied by scholars. Because evolutionary data change over time, new data emerge or old data die out, how to make the data clustering performance at every moment as good as possible, Can basically accurately reflect the data distribution at every moment; discover the evolution mechanism of data through clustering, such as the appearance, change, splitting, disappearing of clustering; and make the clustering results as smooth as possible in time. So that the clustering results at the present time are as similar as possible to those at the previous time, a small number of scholars have studied them. In this paper, we focus on the clustering of evolutionary data, and study two unsupervised evolutionary clustering algorithms and semi-supervised (constrained) evolutionary clustering algorithms, and make a simple application. In this paper, an evolutionary clustering framework based on time smoothness is proposed, which is modified and perfected on the basis of the on-line framework proposed by Chakrabarti et al. In addition, the similarity matrix of data is defined in this paper. The similarity calculation includes the sum of two parts: the similarity between data at current time and the similarity in time series. Finally, the framework is applied to the standard spectral clustering, and two new evolutionary spectral clustering algorithms are obtained and verified by experiments. (2) in this paper, an evolutionary double-layer random walk semi-supervised clustering algorithm is proposed. The algorithm is designed to deal with evolutionary clustering with constraint information. The original static double-layer random walk semi-supervised clustering algorithm spends a lot of time and memory on processing the ever-changing and increasing data, and it can not get a good effect. On the basis of two-layer random walk semi-supervised clustering algorithm, this paper makes good use of the information of the previous time, and directly adds the old data information of the previous moment by solving the similarity between components in the high-level random walk. This paper designs an evolutionary face clustering system, which greatly reduces the computing time and processes the evolution semi-supervised data better, and can get a better clustering result. The face clustering matching in this system is processed by applying the evolutionary clustering algorithm proposed in this paper. The main functions of the system include three parts: face processing (evolutionary clustering, recognition result display, file management).
【学位授予单位】：扬州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13

【相似文献】