Create an SVD/PCA factorization layer

Creates an unconstrained (SVD/PCA) factorization layer. No non-negativity constraint by default; factors are signed.

Usage

svd_layer(
  input,
  k,
  L1 = 0,
  L2 = 0,
  L21 = 0,
  angular = 0,
  upper_bound = 0,
  center = FALSE,
  scale = FALSE,
  method = c("auto", "deflation", "krylov", "lanczos", "irlba", "randomized"),
  mask = NULL,
  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 U and V 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.

center

Center columns before factorization (PCA mode). Default FALSE.

scale

Scale columns to unit variance. Default FALSE.

method

SVD algorithm: "auto", "deflation", "krylov", "lanczos", "irlba", or "randomized". Default "auto".

mask

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

robust

Use robust loss. Default FALSE.

W

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

H

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

name

Optional layer name (for results access).

...

Additional arguments forwarded to svd at fit time. Supports: convergence, cv_seed, patience, k_max, graph_U, graph_V, graph_lambda, threads. See ?svd for the complete list.

Value

An fn_node of type "svd_layer".

Details

SVD-specific parameters (center, scale, method) control the SVD algorithm directly. Regularization parameters and ... are forwarded to svd at fit time.

See also

nmf_layer, factor_input, factor_net, svd

Examples

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

# PCA with centering and scaling
pca_layer <- svd_layer(inp, k = 10, center = TRUE, scale = TRUE)