A description length approach to determining the number of k-means clusters