多种聚类算法性能的比较分析  

Comparative Analysis of Performance of Multiple Clustering Algorithms

作　　者：纪汉霖[1] 李兆信 JI Han-lin;LI Zhao-xin(University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学,上海200093

出　　处：《计算机技术与发展》2020年第8期14-21,共8页Computer Technology and Development

基　　金：国家自然科学基金(71372114);上海市研究生创新基金项目(JWCXSL1102);上海市教委重点学科建设项目(J50504)。

摘　　要：聚类模型是数据挖掘的重要模型,聚类相关模型比较多,聚类算法对数据特征等有较高的要求,目前广泛应用于人工智能、数据分析等领域。选取了六种不同类型的聚类算法,即Affinity Propagation、Birch、Gaussian Mixture Model、Hierarchical clustering、K-means和Spectral,并对其进行了对比分析。采用由机器生成的符合大数据时代下数据特性的数据集而非UCI数据库中符合特定形态的标准测试集,并在数据集上对算法进行了性能测试、效率测试和敏感性分析。研究结果显示,在性能测试上:AP算法聚类效果最佳,其次是K-means算法。Affinity Propagation算法对数据的要求高,使用范围小,K-means适用性和稳定性相对比较好;在效率测试上,Affinity Propagation算法最差,其次是Spectral算法;在敏感性上,K-means算法和Hierarchical clustering算法对数据的数量级不敏感,Spectral算法对数量级比较敏感。从聚类效果、性能和对数量级的敏感性三个方面综合来看,K-means算法相对优于其他五种聚类算法。Clustering model is an important model of data mining.There are many clustering related models and clustering algorithm has high requirements on data characteristics.Currently,it is widely used in artificial intelligence,data analysis and other fields.We have chosen six different types of clustering algorithms including Affinity Propagation,Birch,Gaussian Mixture Model,Hierarchical clustering,K-means and Spectral,which are compared and analyzed.The data set generated by the machine conforms to the data characteristics in the era of big data rather than the standard test set conforming to the specific form in UCI database is adopted,where the algorithm performance test,efficiency test and sensitivity analysis are carried out.The results show that in performance testing,AP has the best clustering effect,followed by K-means.Affinity Propagation has high requirements on data,small range of use and relatively well applicability and stability of K-means.In the aspect of efficiency,Affinity Propagation is the worst,followed by Spectral.In terms of sensitivity,K-means and Hierarchical clustering are not sensitive to the order of magnitude of data,while Spectral is sensitive to the order of magnitude.In terms of clustering effect,performance and sensitivity to order of magnitude,K-means is relatively superior to other five clustering algorithms.

关 键 词：聚类算法 机器数据集 性能 效率 敏感性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]