ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2016, Vol. 53 ›› Issue (5): 1018-1028.doi: 10.7544/issn1000-1239.2016.20150131

• 人工智能 • 上一篇    下一篇



  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) (
  • 出版日期: 2016-05-01
  • 基金资助: 

An Attribute Weighted Clustering Algorithm for Mixed Data Based on Information Entropy

Zhao Xingwang Liang Jiye   

  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: 2016-05-01

摘要: 同时兼具数值型和分类型属性的混合数据在实际应用中普通存在,混合数据的聚类分析越来越受到广泛的关注.为解决高维混合数据聚类中属性加权问题,提出了一种基于信息熵的混合数据属性加权聚类算法,以提升模式发现的效果.工作主要包括:首先为了更加准确客观地度量对象与类之间的差异性,设计了针对混合数据的扩展欧氏距离;然后,在信息熵框架下利用类内信息熵和类间信息熵给出了聚类结果中类内抱团性及一个类与其余类分离度的统一度量机制,并基于此给出了一种属性重要性度量方法,进而设计了一种基于信息熵的属性加权混合数据聚类算法.在10个UCI数据集上的实验结果表明,提出的算法在4种聚类评价指标下优于传统的属性未加权聚类算法和已有的属性加权聚类算法,并通过统计显著性检验表明本文提出算法的聚类结果与已有算法聚类结果具有显著差异性.

关键词: 聚类分析, 混合数据, 属性加权, 信息熵, 相异性度量

Abstract: In real applications, mixed data sets with both numerical attributes and categorical attributes at the same time are more common. Recently, clustering analysis for mixed data has attracted more and more attention. In order to solve the problem of attribute weighting for high-dimensional mixed data, this paper proposes an attribute weighted clustering algorithm for mixed data based on information entropy. The main work includes: an extended Euclidean distance is defined for mixed data, which can be used to measure the difference between the objects and clusters more accurately and objectively. And a generalized mechanism is presented to uniformly assess the compactness and separation of clusters based on within-cluster entropy and between-cluster entropy. Then a measure of the importance of attributes is given based on this mechanism. Furthermore, an attribute weighted clustering algorithm for mixed data based on information entropy is developed. The effectiveness of the proposed algorithm is demonstrated in comparison with the widely used state-of-the-art clustering algorithms for ten real life datasets from UCI. Finally, statistical test is conducted to show the superiority of the results produced by the proposed algorithm.

Key words: clustering analysis, mixed data, attribute weighting, information entropy, dissimilarity measure