Controlling Sparseness in Non-negative Tensor Factorization