Skip to contents

Poisson CP Decomposition (CP-APR) via Multiplicative Updates

Source: R/cp_variants.R
cp_apr.Rd

Fits a CP model to nonnegative (count) data by maximizing the Poisson log-likelihood with the multiplicative-update algorithm of Chi & Kolda, mirroring the MATLAB Tensor Toolbox cp_apr ('mu' method).

Usage

cp_apr(
  X,
  R,
  tol = 1e-04,
  maxiters = 200L,
  maxinner = 10L,
  epsDivZero = 1e-10,
  init = "random",
  printitn = 0L
)

Arguments

X

A nonnegative Tensor or array-like object (typically counts).

R

Target CP rank.

tol

KKT-violation stopping tolerance (default 1e-4).

maxiters

Maximum number of outer iterations (default 200).

maxinner

Maximum inner updates per mode per outer iteration (default 10).

epsDivZero

Safeguard added before divisions (default 1e-10).

init

"random" or a list of initial nonnegative factor matrices.

printitn

Print progress every printitn outer iterations.

Value

A KTensor with nonnegative factors and weights.

References

Chi, E. C. and Kolda, T. G. (2012). On tensors, sparsity, and nonnegative factorizations. SIAM J. Matrix Anal. Appl. 33(4).

Examples

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