Skip to contents

Fit a scaled group lasso regression path

Source: R/sglasso.R
sglasso.Rd

Computes scaled group lasso paths for Gaussian or binary responses.

Usage

sglasso(
  X,
  Y,
  group = 1:ncol(X),
  lambda,
  nlambda = 50,
  d,
  nd = 11,
  alpha = 0.5,
  beta_start = NULL,
  family = "gaussian",
  bilevel = FALSE,
  max_iter = 1e+08,
  eps = 1e-04,
  standardize = TRUE,
  screen = c("SSR", "none", "SSR_fast"),
  diagnostics = FALSE,
  profile = FALSE,
  transform = c("eager", "lazy"),
  lambda.min.ratio = 0.005,
  dfmax = p,
  gmax = length(unique(group)),
  binomial.control = NULL
)

Arguments

X

The design matrix, without an intercept. sglasso standardizes the data and includes an intercept by default.

Y

The response vector.

group

A vector describing the grouping of the coefficients.

lambda

A user supplied sequence of lambda values. Typically, this is left unspecified, and the function automatically computes a grid of lambda values that ranges uniformly on the log scale over the relevant range of lambda values.

nlambda

The number of lambda values. Default is 50

d

The scale parameter between 0 and 1.

nd

The number of d values. Default is 11.

alpha

Elastic Net tuning constant: the value must be between 0 and 1. Default is 0.5.

beta_start

Optional initial coefficient vector.

family

Either "gaussian" (default) or "binomial". Binomial responses must be numeric/logical 0/1 with both classes present.

bilevel

bi-level selection is not supported at this moment.

max_iter

Maximum number of iterations Default is 1e+08.

eps

Convergence threshhold. The algorithm iterates until the BCD for the change in linear predictors for each coefficient is less than eps. Default is 1e-4.

standardize

Logical flag indicating whether X should be centered and scaled internally before fitting. The default is TRUE, preserving the usual package behavior. Set to FALSE only when X has already been centered/scaled using the intended training-data transformation.

screen

Screening rule used before the final KKT checks. One of "SSR", "none", or "SSR_fast". The "SSR_fast" mode checks the rest-set KKT conditions at the first, first-quartile, median, third-quartile, and final lambda values. The default is "SSR".

diagnostics

Logical flag indicating whether detailed per-lambda timing and final inactive-set KKT residual diagnostics should be computed. The default is FALSE to avoid diagnostic overhead in ordinary fits.

profile

Logical flag indicating whether coarse R-level runtime components should be recorded. The default is FALSE.

transform

Coefficient transformation mode. The default "eager" returns the fitted coefficients on the original data scale during sglasso(). The experimental "lazy" mode stores the solver-scale path and performs the back-transformation when coef() is called, which can reduce fit-time overhead in runtime comparisons.

lambda.min.ratio

The smallest value for lambda, as a fraction of the starting reference. Default is .005. For binomial models this is a finite search range, not a guarantee of a null model; see sglasso-binomial.

dfmax

Limit on the number of parameters allowed to be nonzero. If this limit is exceeded, the algorithm will exit early from the regularization path.

gmax

Limit on the number of groups allowed to have nonzero elements. If this limit is exceeded, the algorithm will exit early from the regularization path.

binomial.control

Named list of numerical controls for the binomial solver; see sglasso-binomial. Ignored for Gaussian fits.

Value

An object with S3 class "sglasso" containing:

beta

The fitted matrix of coefficients. The number of rows is equal to the number of coefficients, and the number of columns is equal to nlambda.

family

Same as above.

group

Same as above.

lambda

The sequence of lambda values in the path.

alpha

Same as above.

deviance

A vector containing the deviance of the fitted model at each value of `lambda`.

n

Number of observations.

penalty

Same as above.

df

A vector of length `nlambda` containing estimates of effective number of model parameters all the points along the regularization path. For details on how this is calculated, see Breheny and Huang (2009).

iter

A vector of length `nlambda` containing the number of iterations until convergence at each value of `lambda`.

group.multiplier

A named vector containing the multiplicative constant applied to each group's penalty.

See also

cv.sglasso

Author

Bahadir Yuzbasi and Jiguo Cao

Examples

data(GenAtHum,package = "sglasso")
X <- GenAtHum$X
y <- GenAtHum$y
group <- GenAtHum$group
n =  nrow(X)
p =  ncol(X)
set.seed(123)
fit <- sglasso(X = X, Y = y, group = group, nlambda = 20, nd = 3)
select(fit,"EBIC")