Ensemble-Initialized k-Means Clustering

As one of the most classical clustering techniques, the k-means clustering has been widely used in various areas over the past few decades. Despite its significant success, there are still several challenging issues in the k-means clustering research, one of which lies in its high sensitivity to the selection of the initial cluster centers. In this paper, we propose a new cluster center initialization method for k-means based on ensemble learning. Specifically, an ensemble of base clusterings are first constructed by using multiple k-means clusterers with random initializations. Then, a co-association matrix is computed for the base clusterings, upon which the agglomerative clustering algorithm can thereby be performed to build a pre-clustering result. From the pre-clustering, the set of initial cluster centers are obtained and then used for the final k-means clustering process. Experiments on multiple real world datasets have demonstrated the superiority of the proposed method.