cv.starnet {starnet}

Model comparison

Description

Compares stacked elastic net, tuned elastic net, ridge and lasso.

Usage

cv.starnet(
  y,
  X,
  family = "gaussian",
  nalpha = 21,
  alpha = NULL,
  nfolds.ext = 10,
  nfolds.int = 10,
  foldid.ext = NULL,
  foldid.int = NULL,
  type.measure = "deviance",
  alpha.meta = 1,
  nzero = NULL,
  intercept = NULL,
  upper.limit = NULL,
  unit.sum = NULL,
  ...
)

Arguments

y

response: numeric vector of length n

X

covariates: numeric matrix with n rows (samples) and p columns (variables)

family

character "gaussian", "binomial" or "poisson"

nalpha

number of alpha values

alpha

elastic net mixing parameters: vector of length nalpha with entries between 0 (ridge) and 1 (lasso); or NULL (equidistance)

nfolds.ext, nfolds.int, foldid.ext, foldid.int

number of folds (nfolds): positive integer; fold identifiers (foldid): vector of length n with entries between 1 and nfolds, or NULL, for hold-out (single split) instead of cross-validation (multiple splits): set foldid.ext to 0 for training and to 1 for testing samples

type.measure

loss function: character "deviance", "class", "mse" or "mae" (see cv.glmnet)

alpha.meta

meta-learner: value between 0 (ridge) and 1 (lasso) for elastic net regularisation; NA for convex combination

nzero

number of non-zero coefficients: scalar/vector including positive integer(s) or NA; or NULL (no post hoc feature selection)

intercept

settings for meta-learner: logical, or NULL (intercept=!is.na(alpha.meta), upper.limit=TRUE, unit.sum=is.na(alpha.meta))

upper.limit

settings for meta-learner: logical, or NULL (intercept=!is.na(alpha.meta), upper.limit=TRUE, unit.sum=is.na(alpha.meta))

unit.sum

settings for meta-learner: logical, or NULL (intercept=!is.na(alpha.meta), upper.limit=TRUE, unit.sum=is.na(alpha.meta))

...

further arguments (not applicable)

Value

List containing the cross-validated loss (or out-of sample loss if nfolds.ext equals two, and foldid.ext contains zeros and ones). The slot meta contains the loss from the stacked elastic net (stack), the tuned elastic net (tune), ridge, lasso, and the intercept-only model (none). The slot base contains the loss from the base learners. And the slot extra contains the loss from the restricted stacked elastic net (stack), lasso, and lasso-like elastic net (enet), with the maximum number of non-zero coefficients shown in the column name.

Examples

loss <- cv.starnet(y=y,X=X)