Analytical and numerical evaluation of the suppressed fuzzy <i>c</i>-means algorithm: a study on the competition in <i>c</i>-means clustering models

作     者：Szilagyi, Laszlo Szilagyi, Sandor M. Benyo, Zoltan 

作者机构：Sapientia Univ Transylvania Fac Tech & Human Sci Corunca 547367 Romania Budapest Univ Technol & Econ Dept Control Engn & Informat Technol H-1117 Budapest Hungary 

出 版 物：《SOFT COMPUTING》 (Soft Comput.)

年 卷 期：2010年第14卷第5期

页      面：495-505页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Communitas and Eurotrans Foundations of Transylvania Hungarian National Office for Research and Technology Sapientia Institute for Research Programmes 

主　　题：Fuzzy c-means algorithm Suppressed fuzzy c-means algorithm Competitive clustering Alternating optimization Learning rate 

摘      要：Suppressed fuzzy c-means (s-FCM) clustering was introduced in Fan et al. (Pattern Recogn Lett 24: 1607-1612, 2003) with the intention of combining the higher speed of hard c-means (HCM) clustering with the better classification properties of fuzzy c-means (FCM) algorithm. The authors modified the FCM iteration to create a competition among clusters: lower degrees of memberships were diminished according to a previously set suppression rate, while the largest fuzzy membership grew by swallowing all the suppressed parts of the small ones. Suppressing the FCM algorithm was found successful in the terms of accuracy and working time, but the authors failed to answer a series of important questions. In this paper, we clarify the view upon the optimality and the competitive behavior of s-FCM via analytical computations and numerical analysis. A quasi competitive learning rate (QLR) is introduced first, in order to quantify the effect of suppression. As the investigation of s-FCM s optimality did not provide a precise result, an alternative, optimally suppressed FCM (Os-FCM) algorithm is proposed as a hybridization of FCM and HCM. Both the suppressed and optimally suppressed FCM algorithms underwent the same analytical and numerical evaluations, their properties were analyzed using the QLR. We found the newly introduced Os-FCM algorithm quicker than s-FCM at any nontrivial suppression level. Os-FCM should also be favored because of its guaranteed optimality.