Rethinking k-means from manifold learning perspective