An approach to non-negative matrix factorization that does not require data to be separable and provides a generally unique decomposition.