Composite likelihood inference in a discrete latent variable model for two-way “clustering-by-segmentation” problems