parafac2(tensor_slices, rank, n_iter_max=100, init='random', svd='numpy_svd', normalize_factors=False, tol=1e-08, random_state=None, verbose=False, return_errors=False, n_iter_parafac=5)[source]

PARAFAC2 decomposition [1] of a third order tensor via alternating least squares (ALS)

Computes a rank-rank PARAFAC2 decomposition of the third-order tensor defined by tensor_slices. The decomposition is on the form (A [B_i] C) such that the

i-th frontal slice, X_i, of X is given by

X_i = B_i diag(a_i) C^T,

where diag(a_i) is the diagonal matrix whose nonzero entries are equal to the i-th row of the I \times R factor matrix A, B_i is a J_i \times R factor matrix such that the cross product matrix B_{i_1}^T B_{i_1} is constant for all i, and C is a K \times R factor matrix. To compute this decomposition, we reformulate the expression for B_i such that

B_i = P_i B,

where P_i is a J_i \times R orthogonal matrix and B is a R \times R matrix.

An alternative formulation of the PARAFAC2 decomposition is that the tensor element X_{ijk} is given by

X_{ijk} = \sum_{r=1}^R A_{ir} B_{ijr} C_{kr},

with the same constraints hold for B_i as above.

tensor_slicesndarray or list of ndarrays

Either a third order tensor or a list of second order tensors that may have different number of rows. Note that the second mode factor matrices are allowed to change over the first mode, not the third mode as some other implementations use (see note below).


Number of components.


Maximum number of iteration

init{‘svd’, ‘random’, KruskalTensor, Parafac2Tensor}

Type of factor matrix initialization. See initialize_factors.

svdstr, default is ‘numpy_svd’

function to use to compute the SVD, acceptable values in tensorly.SVD_FUNS

normalize_factorsbool (optional)

If True, aggregate the weights of each factor in a 1D-tensor of shape (rank, ), which will contain the norms of the factors. Note that there may be some inaccuracies in the component weights.

tolfloat, optional

(Default: 1e-8) Relative reconstruction error tolerance. The algorithm is considered to have found the global minimum when the reconstruction error is less than tol.

random_state{None, int, np.random.RandomState}
verboseint, optional

Level of verbosity

return_errorsbool, optional

Activate return of iteration errors

n_iter_parafac: int, optional

Number of PARAFAC iterations to perform for each PARAFAC2 iteration

Parafac2Tensor(weight, factors, projection_matrices)
  • weights1D array of shape (rank, )

    all ones if normalize_factors is False (default), weights of the (normalized) factors otherwise

  • factorsList of factors of the CP decomposition element i is of shape

    (tensor.shape[i], rank)

  • projection_matricesList of projection matrices used to create evolving



A list of reconstruction errors at each iteration of the algorithms.


This formulation of the PARAFAC2 decomposition is slightly different from the one in [1]. The difference lies in that here, the second mode changes over the first mode, whereas in [1], the second mode changes over the third mode. We made this change since that means that the function accept both lists of matrices and a single nd-array as input without any reordering of the modes.


[1](1, 2, 3) Kiers, H.A.L., ten Berge, J.M.F. and Bro, R. (1999), PARAFAC2—Part I. A direct fitting algorithm for the PARAFAC2 model. J. Chemometrics, 13: 275-294.