Fast Convegence Clustering Ensemble

Clustering ensemble combines some clustering outputs to obtain better results. High robustness, accuracy and stability are the most important characteristics of clustering ensembles. Previous clustering ensembles usually use k-means to generate ensemble members. The main problem of k-means is initial samples which have high effect on final results. Refining initial samples of kmeans increases the complexity of algorithm significantly. In this paper we try to predict initial samples, especially for clustering ensemble, without any increasing in time complexity. In this paper we introduce two approaches to select the initial samples of k-means intelligently to generate ensemble members. The proposed methods increase both accuracy and the speed of convergence without any increasing in time complexity. Selecting one sample from each cluster of previous result and selecting k samples which have minimum similarity to each other from coassociation matrix are the two proposed method in refining initial samples of k-means. Clarity, simplicity, fast convergence and higher accuracy are the most important parameters of proposed algorithm. Experimental results demonstrate the effect of proposed algorithm in convergence and accuracy of common datasets.