CP Alternating Least Squares Decomposition

Computes a rank-R canonical polyadic (CP) decomposition of a dense tensor by alternating least squares, mirroring the behavior of the MATLAB Tensor Toolbox cp_als.

Usage

cp_als(
  X,
  R,
  tol = 1e-04,
  maxiters = 50L,
  dimorder = NULL,
  init = "random",
  printitn = 0L,
  fixsigns = TRUE
)

Arguments

X: A Tensor or array-like object.
R: Target CP rank (positive integer).
tol: Convergence tolerance on change in fit (default 1e-4).
maxiters: Maximum number of ALS sweeps (default 50).
dimorder: Integer permutation of 1:ndims(X) giving the order in which factor matrices are updated. Defaults to 1:ndims(X).
init: Either "random" (i.i.d. normal), "nvecs" (leading left singular vectors of the mode-n unfolding; falls back to random for modes where the mode size is smaller than R), or a list of initial factor matrices.
printitn: Print fit every printitn iterations (0 to suppress).
fixsigns: Logical; if TRUE, resolve sign ambiguity of the returned components.

Value

A KTensor giving the rank-R CP decomposition of X.

Examples

set.seed(1)
X <- tensor(array(runif(24), dim = c(2, 3, 4)))
K <- cp_als(X, R = 2, maxiters = 20)