层次聚类算法的改进及分析.pdfVIP

第 25 卷第 6期 计算机应用与软件 V o l125 No. 6 2008年 6 月 Com puter A pplications and Sofwt are Jun. 2008 1 1 2 郭晓娟 刘晓霞 李晓玲 1 ( 7 10 127 2 ( 430074 层次凝聚算法是一 个非常有用的聚类算法, 它在迭代地凝聚每次接近对 到所有的数据都属于同一 个簇但层次聚类 也存在着几 个缺点, 如聚类时的时空复杂性高; 聚类的簇效率低误差较大等经验研究表明, 大部分 HA C 算法都有这样一 个趋势: 除了在谱系图的顶层, 所有低层聚类的簇都是比较小的并且很接近于其他的簇, 提出了一种改进算法能够减小时空复杂性并能验证 其正确性, 分析与实验都证明这种方法是非常有效的 聚类 层次聚类 谱系图 簇 POP ON IMPROVEMENT AND ANALYSIS OF H IERARCHICAL CLUSTERING ALGOR ITHM 1 1 2 Guo X iaojuan L iu X iaox ia L i X iaoling 1 ( N orthw est University , X ipan 710127, Shaanxi, Ch ina) 2 (Ch ina University of Geosciences, Wuhan4 0074, H ube,i China) Abstract A prom inent and u se fu l c lass of algor ithm is h ie rarch ica l agg lom erative clu ster ing ( HAC w h ich iteratively agg lom erates the c lo- sest pare until a ll data po ints be long to one c lu ste r. How ever, HAC m ethods have several draw backs, such as high tmi e andm emo ry com plex- ities w hen c lu ster ing, insuff ic ient and inaccura te clu ster va lida tion, etc. Em p ir ica l study show s that mo st HA C a lgor ithm s fo llow a trend w here, except fo r a number of top leve ls o f the dendrogram, all low er leve l agg lom erate c lu sters are very sm a ll in size and close in prox mi ity to other c lu sters. M ethods a re proposed to reduce the tmi e and m em ory comp lex ities sign ificantly and to m ake validation v ery e ff ic ient and ac- curate. A nalysis and exper mi ents a ll prove the effectiveness o f the proposed m ethod. K eywords C lu stering HA C D endrogram C lu ste r POP