Skip to contents

Create an NMF factorization layer

Source: R/factor_net.R
nmf_layer.Rd

Creates a non-negative matrix factorization layer that decomposes its input into W * diag(d) * H with non-negativity constraints.

Usage

nmf_layer(
  input,
  k,
  L1 = 0,
  L2 = 0,
  L21 = 0,
  angular = 0,
  upper_bound = 0,
  mask = NULL,
  zi = c("none", "row", "col"),
  projective = FALSE,
  symmetric = FALSE,
  robust = FALSE,
  W = NULL,
  H = NULL,
  name = NULL,
  ...
)

Arguments

input

An fn_node (input, shared, condition, or another layer).

k

Factorization rank.

L1

L1 penalty (shared by W and H unless overridden). Default 0.

L2

L2 penalty. Default 0.

L21

Group sparsity penalty. Default 0.

angular

Orthogonality penalty. Default 0.

upper_bound

Box constraint. Default 0.

mask

Masking mode: NULL (none), "zeros", "NA", or a sparse mask matrix. See nmf for details.

zi

Zero-inflation mode: "none", "row", or "col". Requires loss = "gp" or "nb" in factor_config(). Default "none".

projective

Use projective NMF (W is reused as H). Default FALSE.

symmetric

Use symmetric NMF (W == H). Default FALSE.

robust

Robustness control: FALSE (default, no robustness), TRUE (Huber delta=1.345), "mae" (near-MAE, delta=1e-4), or a positive numeric Huber delta. See nmf for details.

W

Optional W() config to override W-specific settings.

H

Optional H() config to override H-specific settings.

name

Optional layer name (for results access).

...

Additional arguments forwarded to nmf at fit time. Supports all advanced parameters: distribution tuning (dispersion, theta_init, etc.), IRLS control (irls_max_iter, irls_tol), solver tuning (cd_tol, cd_maxit), streaming (streaming, panel_cols), callbacks (on_iteration), and more. See ?nmf for the complete list.

Value

An fn_node of type "nmf_layer".

Details

When used with |>, the input node is the first argument: x |> nmf_layer(k = 64).

Layer-level regularization parameters (L1, L2, etc.) apply to both W and H unless overridden by W() or H().

See also

svd_layer, factor_net, factor_input, nmf

Examples

data(aml)
inp <- factor_input(aml)
layer <- nmf_layer(inp, k = 5)

# With zero-inflation and distribution-specific tuning
layer_gp <- nmf_layer(inp, k = 5, zi = "row",
                      dispersion = "per_col", theta_init = 0.5)