Skip to contents

Global configuration for a factorization network

Source: R/factor_net.R
factor_config.Rd

Sets network-wide defaults. Layer-level and factor-level settings override these where specified.

Usage

factor_config(
  maxit = 100,
  tol = 1e-04,
  loss = c("mse", "gp", "nb", "gamma", "inverse_gaussian", "tweedie"),
  verbose = FALSE,
  seed = NULL,
  threads = 0,
  norm = c("L1", "L2", "none"),
  solver = c("auto", "cholesky", "cd"),
  resource = "auto",
  test_fraction = 0,
  cv_seed = 0L,
  mask_zeros = FALSE,
  patience = 5L,
  ...
)

Arguments

maxit

Maximum ALS iterations per layer. Default 100.

tol

Convergence tolerance. Default 1e-4.

loss

Loss function: "mse", "gp", "nb", "gamma", "inverse_gaussian", "tweedie". Default "mse". For robust (Huber/MAE) fitting, use the robust parameter at the layer level instead.

verbose

Print per-iteration diagnostics. Default FALSE.

seed

Random seed. Default NULL (auto).

threads

Number of CPU threads (0 = all). Default 0.

norm

Normalization type: "L1", "L2", "none". Default "L1".

solver

NNLS solver: "auto", "cholesky", "cd". Default "auto".

resource

Resource override: "auto", "cpu", "gpu". Default "auto".

test_fraction

Fraction of entries held out for cross-validation test set. 0 = no CV (standard fit). Default 0.

cv_seed

Separate seed for CV mask (0 = derive from seed).

mask_zeros

If TRUE, only non-zero entries are in the test set (suitable for recommendation systems). If FALSE, all entries including zeros may appear in the test set. Default FALSE.

patience

Number of iterations without test-loss improvement before early stopping during CV. Default 5.

...

Additional arguments applied network-wide. These are forwarded to the underlying nmf or svd call at fit time as lowest-priority defaults (layer-level ... overrides these). Supports distribution tuning (dispersion, theta_init, etc.), IRLS control, solver tuning, and more. See ?nmf or ?svd for the complete list.

Value

An fn_global_config object.

Details

Unsupported combinations:

  • Cholesky solver with non-MSE losses (GP, NB, Gamma, etc.) is not supported. Use solver="cd" or solver="auto".

  • Zero-inflation is only available with GP and NB losses at the layer level; it cannot be set globally in factor_config().

See also

factor_net, nmf_layer, fit

Examples

# Default config
cfg <- factor_config()

# Poisson NMF with cross-validation
cfg <- factor_config(loss = "gp", test_fraction = 0.1, maxit = 50)