An initialization method for the K-Means algorithm using neighborhood model

As a simple clustering method, the traditional K-Means algorithm has been widely discussed and applied in pattern recognition and machine learning. However, the K-Means algorithm could not guarantee unique clustering result because initial cluster centers are chosen randomly. In this paper, the cohesion degree of the neighborhood of an object and the coupling degree between neighborhoods of objects are defined based on the neighborhood-based rough set model. Furthermore, a new initialization method is proposed, and the corresponding time complexity is analyzed as well. We study the influence of the three norms on clustering, and compare the clustering results of the K-means with the three different initialization methods. The experimental results illustrate the effectiveness of the proposed method. (C) 2009 Elsevier Ltd. All rights reserved.