MsSparse: penalized sparse NMF engine specification

[MsSparse] is the concrete [MsEngine] spec for penalized sparse NMF extraction (U-M2-05): `"volume"` = KL likelihood with a log-det volume regularizer on W (Leplat-Gillis-Ang), `"l1"` = KL with L1 penalties on both factors (Hoyer/Cichocki KL-L1L1 MU; SparseSignatures itself is least-squares + LASSO — attribution erratum in engine/sparse.rs). mu = 0 degenerates). Build one with [ms_sparse()]. Violations raise `msuiter_error_engine`.

Usage

MsSparse(
  name = "sparse",
  mode = "extract",
  params = list(),
  deterministic = TRUE,
  packages = character(0),
  label = "penalized sparse NMF",
  variant = "l1",
  lambda = 0,
  mu = 0,
  delta = 0,
  tol = 1e-07,
  max_iter = 2000,
  seed = 1
)

Arguments

name

registry key; fixed to `"sparse"`.

mode

spec mode; fixed to `"extract"`.

params

named list of hyper-parameters (empty: the sparse hyper-parameters live in the dedicated properties below).

deterministic

always TRUE: the kernel is seeded and deterministic.

packages

character vector of required R packages; none.

label

human-readable engine label.

variant

`"volume"` or `"l1"`.

lambda

penalty on the signatures (volume strength / H-row L1).

mu

W-column L1 penalty (l1 variant only; must be > 0 there — mu = 0 degenerates the exposure scale).

delta

volume regularizer ridge (volume variant only; 0 picks the kernel default 1).

tol

convergence tolerance on the relative objective change.

max_iter

single positive whole number of update iterations.

seed

single whole number in `[0, 2^31 - 1]`.