Biclustering

Biclustering is a family of methods that operate on matrices whose rows and columns represent different features connected with a precise rationale. For example, the rows can represent customers, and the columns products. Each element can indicate a rating or, if zero, the fact that a specific product, p_j, has not been bought/rated by the customer, c_i. As the behavior of the customers can generally be segmented into specific sets, we can assume that A has an underlying checkerboard structure, where the compact regions, called biclusters, represent sub-matrices with peculiar properties.

The nature of such properties depends on the specific context, but the structures share the common feature of being strongly separated from the remaining regions. In our example, the biclusters can be mixed segments containing sets of customers and products that agree on the rating (this concept will be clearer in the practical example), but more generally, the rearrangement of rows...