A Robust Clustering Approach Based on KNN and Modified C-Means Algorithm
Amir Aliabadian, Department of Faculty Member of Electrical and Computer Engineering, Shomal University, Iran.
Manuscript received on March 27, 2014. | Revised Manuscript received on April 09, 2014. | Manuscript published on April 30, 2014. | PP:128-132 | Volume-3, Issue-4, April 2014. | Retrieval Number: F2072082613/2013©BEIESP
Open Access | Ethics and Policies | Cite
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Cluster analysis is used for clustering a data set into groups of similar individuals. It is an approach towards to unsupervised learning and is one of the major techniques in pattern recognition. FCM algorithm needs the number of classes and initial values of center for each cluster. These values are determined randomly, so it may cause target function converges to several local center. so many iterative stages are needed, until FCM can reach to global center for each cluster. In this paper, we suggest robust hybrid algorithm in which, we have real unsupervised learning algorithm, no need to initial center value and the number of clusters. The First layer in this algorithm finds initial clustering center by K-nearest neighbor (K-NN) rules based on unsupervised learning approach. In the second layer, we applied FCM only one time for having optimal clustering. It is done by means of Fuzzy clustering validation criterion, unlike FCM that needs iterative process. We applied new algorithm to several set of standard databases (IRIS).results show that this algorithm is more accurate than FCM both in estimation of optimal number of clusters and correctness of devotion of data to their real clusters.
Keywords: Cluster analysis. FCM algorithm. K-nearest neighbor, Target function.