Limitations of k-means clustering