一种基于投票策略的聚类融合算法  

A Clustering Ensembles Algorithm Based on Voting Strategy

作　　者：李金磊[1] 朱晓莲[2] 朱海燕[2] 

机构地区：[1]中国石化勘探南方分公司研究院,四川成都610041 [2]中国地质大学(武汉)计算机学院,湖北武汉430074

出　　处：《计算机仿真》2008年第3期126-128,133,共4页Computer Simulation

摘　　要：在分类算法和回归模型中,融合方法正得到越来越广泛的应用,但在非监督机器学习领域,由于缺乏数据集的先验知识,则不能直接用于聚类算法。提出并实现了一种基于投票策略的聚类融合算法,该算法利用k-means算法每次随机选取聚类中心而得到不同样本划分的特性,将多次运行得到的聚类结果通过投票的方式合并,从而得到最终的结果。通过一系列真实数据和合成数据集的实验证明,这种方法比单一的聚类算法能更有效地提高聚类的准确率。在此基础上,为了降低高维数据运算的复杂性,将随机划分属性子空间的方法应用到上述聚类融合算法中,实验证明,该方法同时也能够在一个属性子空间上获得好的聚类结果。In the classification and regression algorithms, the ensemble method was widely used, but in the unsupervised learning, it didn＇ t be used in the clustering algorithm directly due to lack of prior knowledge. This paper proposed a clustering ensembles algorithm based on voting strategy, it used the characteristic that the k - means algorithm selected the clustering centers randomly and found the different partitions of the sample. Then, it combined the clustering results of operating the k - means algorithm repeatedly to a final result via the voting strategy. Through the experiment on lots of real data and artificial data, this method could receive a better result than the single clustering algorithm. Moreover, to resolve the complexity of the high dimension data, a method for partitioning the feature space randomly in the ensemble algorithm was used. It proved that this method is able to receive a good clustering result in the attribute subspace by the experiment.

关 键 词：聚类融合 均值算法 投票策略 属性子空间 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]