Clustering by latent dimensions