ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2016, Vol. 53 ›› Issue (9): 1979-1989.doi: 10.7544/issn1000-1239.2016.20150592

• 人工智能 • 上一篇    下一篇

一种不完备混合数据集成聚类算法

史倩玉,梁吉业,赵兴旺   

  1. (山西大学计算机与信息技术学院 太原 030006) (计算智能与中文信息处理教育部重点实验室(山西大学) 太原 030006) (ljy@sxu.edu.cn)
  • 出版日期: 2016-09-01
  • 基金资助: 
    国家自然科学基金重点项目(61432011);国家自然科学基金项目(61573229,61502289);山西省科技基础条件平台建设项目(2012091002-0101);山西省自然科学基金项目(201601D202039);山西省研究生教育创新项目(2016SY002)

A Clustering Ensemble Algorithm for Incomplete Mixed Data

Shi Qianyu, Liang Jiye, Zhao Xingwang   

  1. (School of Computer and Information Technology, Shanxi University, Taiyuan 030006) (Key Laboratory of Computational Intelligence and Chinese Information Processing(Shanxi University), Ministry of Education, Taiyuan 030006)
  • Online: 2016-09-01

摘要: 集成聚类技术由于具有较好的泛化能力,目前引起了研究者的高度关注.已有研究主要关注数值型完备数据的集成聚类问题.然而,实际应用中面临的数据往往是兼具数值属性和分类属性共同描述的混合型数据,而且通常带有缺失值.为此,针对不完备混合数据提出了一种集成聚类算法,首先利用3种缺失值填充方法对不完备混合数据进行完备化处理;其次在3种填充后的不同完备数据集上分别多次执行K-Prototypes算法产生基聚类结果;最后对基聚类结果进行集成.在UCI真实数据集上与传统聚类算法通过实验进行了比较分析,实验结果表明提出的算法是有效的.

关键词: 集成聚类, 不完备数据, 混合数据, 缺失值填充, K原型聚类算法

Abstract: Cluster ensembles have recently emerged a powerful clustering analysis technology and caught high attention of researchers due to their good generalization ability. From the existing work, these techniques held great promise, most of which generate the final results for complete data sets with numerical attributes. However, real life data sets are usually incomplete mixed data described by numerical and categorical attributes at the same time. And these existing algorithms are not very effective for an incomplete mixed data set. To overcome this deficiency, this paper proposes a new clustering ensemble algorithm which can be used to ensemble final clustering results for mixed numerical and categorical incomplete data. Firstly, the algorithm conducts completion of incomplete mixed data using three different missing value filling methods. Then, a set of clustering solutions are produced by executing K-Prototypes clustering algorithm on three different kinds of complete data sets multiple times, respectively. Next, a similarity matrix is constructed by considering all the clustering solutions. After that, the final clustering result is obtained by hierarchical clustering algorithms based on the similarity matrix. The effectiveness of the proposed algorithm is empirically demonstrated over some UCI real data sets and three benchmark evaluation measures. The experimental results show that the proposed algorithm is able to generate higher clustering quality in comparison with several traditional clustering algorithms.

Key words: clustering ensemble, incomplete data, mixed data, missing value imputation, K-Prototypes clustering algorithm

中图分类号: