A new algorithm for choosing initial cluster centers for k-means

The k-means algorithm is widely used in many applications due to its simplicity and fast speed. However, its result is very sensitive to the initialization step: choosing initial cluster centers. Different initialization algorithms may lead to different clustering results and may also affect the convergence of the method. In this paper, we propose a new algorithm for improving the initialization of the cluster centers by reducing dimensions followed by moving cluster centers towards high density regions. Our algorithm is compared with three other initialization algorithms for k-means. And the effectiveness of our approach is shown by a series of carefully designed experiments.