Cross-validation for scaled group lasso

Performs K-fold cross-validation for the scaled group lasso over a grid of lambda and d values.

Usage

cv.sglasso(
  X,
  Y,
  group = 1:ncol(X),
  lambda,
  nlambda = 100,
  d,
  nd = 11,
  alpha = 0.5,
  nfolds = 10,
  fold,
  beta_start = NULL,
  family = "gaussian",
  bilevel = FALSE,
  max_iter = 1e+08,
  eps = 1e-04,
  standardize = TRUE,
  screen = c("SSR", "none", "SSR_fast"),
  ...
)

Arguments

X

Design matrix.

Y

Response vector.

group

Group membership vector.

lambda

Optional lambda sequence.

nlambda

Number of lambda values.

d

Optional d sequence.

nd

Number of d values.

alpha

Mixing parameter.

nfolds

Number of cross-validation folds.

fold

Optional fold assignment vector.

beta_start

Optional initial coefficient vector.

family

Model family: "gaussian" or "binomial".

bilevel

Logical; whether bilevel selection is used.

max_iter

Maximum number of iterations.

eps

Convergence tolerance.

standardize

Logical flag indicating whether X should be centered and scaled internally before fitting.

screen

Screening rule passed to sglasso.

...

For binomial fits: lambda.min.ratio and binomial.control, passed to the numerical adapter.

Value

An object of class cv.sglasso.

Details

Binomial CV minimizes mean out-of-fold log-loss, using raw training rows to recompute standardization, group orthonormalization and Firth targets separately in every fold. Automatic paths align by a common relative grid with fold-specific training-only reference scales; explicit lambda aligns by absolute values. alpha is a fixed scalar, not tuned by this function. Use the same supplied folds for comparisons across alpha values. The test set must not be supplied to this function. See sglasso-binomial for finite-range and numerical limitations.

See also

sglasso

Examples

data(Birthwt, package = "grpreg")
X <- Birthwt$X
group <- Birthwt$group
Y <- Birthwt$bwt
CVsglasso_fit <- cv.sglasso(X, Y, group)