First release since v4.1-10 (July 2025). Major rearchitecture with Cox speedups, a new per-call algorithm-control mechanism, and two new vignettes.
cox.ties = c("breslow", "efron");
default is "breslow" for now. Calls with family = "cox" that
don't set cox.ties emit a warning — the default switches to
"efron" in v5.1 to match survival::coxph.control = list(...) argument to glmnet() (and to
cv.glmnet, glmnet.path, glmnet.fit, elnet.fit) overrides
any of 17 algorithm-control parameters for a single call without
mutating session state. Unknown keys error.thresh, maxit, dfmax, pmax, trace.it in
glmnet() are deprecated; use control = list(...) or
glmnet.control().dfmax and pmax signature defaults are now NULL, resolving at
call time to nvars + 1 and min(2*dfmax + 20, nvars).glmnet-history.Rmd — package history, milestones,
attributions.coxdev.Rmd — the coxdev deviance/gradient library.predict.glmnet(type = "nonzero", s = ...) now always returns a
list (one slot per s value, each slot an integer vector of
active-coefficient indices). Earlier versions returned a
data.frame when the per-s index counts happened to be uniform,
and a list otherwise. This makes the return shape consistent with
the no-s form and with the function's documentation, but is a
silent breaking change for callers that branched on
is.data.frame() of the result. To get a stable union of indices
across s values regardless of glmnet version, use
sort(unique(unlist(predict(fit, type = "nonzero", s = lambda)))).Adjusted plotting again. Default is now using -log(lambda) as default for plotting, which is option xvar='lambda'. Also fixed labeling issues with predict and coef
Changed default plotting xvar = "lambda". Corrected some minor issues in citation file.
Added DOI for JSS 2023 paper and corrected some typos in documentation
(nfold -> nfolds) and vignette.
Removed unneeded legacy fortran code, leaving only coxnet. Fixed up Matrix as() sequences
Relatively minor changes to bugs in survival functions and bigGlm, and some improved failure messages.
Most of the Fortran code has been replaced by C++ by James Yang, leading to speedups in all cases. The exception is the Cox routine for right censored data, which is still under development.
Some of the Fortran in glmnet has been replaced by C++, written by the newest member of our team, James Yang.
wls routines (dense and sparse), that are the engines under the glmnet.path
function when we use programmable families, are now written in C++, and lead
to speedups of around 8x.glmnet(...,family="gaussian") are all in C++, and lead to speedups
around 4x.A new feature added, as well as some minor fixes to documentation.
newx observation failed before. This is
fixed.cv.glmnet improved.Fixed some bugs in the coxpath function to do with sparse X.
cv.glmnet to explain implications of
supplying lambdaExpanded scope for the Cox model.
survival::coxphsurvival::survfitVignettes are revised and reorganized.
Additional index information stored on cv.glmnet objects, and
included when printed.
concordance
function from package survivalMajor revision with added functionality. Any GLM family can be used
now with glmnet, not just the built-in families. By passing a
"family" object as the family argument (rather than a character
string), one gets access to all families supported by glm. This
development was programmed by our newest member of the glmnet team,
Kenneth Tay.
Bug fixes
Intercept=FALSE with "Gaussian" is fixed. The dev.ratio comes out
correctly now. The mortran code was changed directly in 4
places. look for "standard". Thanks to Kenneth Tay.Bug fixes
confusion.glmnet was sometimes not returning a list because of
apply collapsing structurecv.mrelnet and cv.multnet dropping dimensions inappropriatelystorePB to avoid segfault. Thanks Tomas Kalibera!assess.glmnet and cousins to be more helpful!lambda.interp to avoid edge cases (thanks
David Keplinger)Minor fix to correct Depends in the DESCRIPTION to R (>= 3.6.0)
This is a major revision with much added functionality, listed
roughly in order of importance. An additional vignette called relax
is supplied to describe the usage.
relax argument added to glmnet. This causes the
models in the path to be refit without regularization. The
resulting object inherits from class glmnet, and has an
additional component, itself a glmnet object, which is the relaxed
fit.relax argument to cv.glmnet. This allows selection from a
mixture of the relaxed fit and the regular fit. The mixture is
governed by an argument gamma with a default of 5 values between
0 and 1.predict, coef and plot methods for relaxed and cv.relaxed
objects.print method for relaxed object, and new print methods for
cv.glmnet and cv.relaxed objects.trace.it=TRUE to glmnet and cv.glmnet. This can also be set
for the session via glmnet.control.assess.glmnet, roc.glmnet and
confusion.glmnet for displaying the performance of models.makeX for building the x matrix for input to glmnet. Main
functionality is one-hot-encoding of factor variables, treatment
of NA and creating sparse inputs.bigGlm for fitting the GLMs of glmnet unpenalized.In addition to these new features, some of the code in glmnet has
been tidied up, especially related to CV.
coxnet.deviance to do with input
pred, as well as saturated loglike (missing) and weightscoxgrad function for computing the gradientcv.glmnet, for cases when wierd
things happeninst/mortraninst/mortran-Wall warningsnewoffset created problems all over - fixed theseexact=TRUE calls to coef and
predict. See help file for more detailsy blows up elnet; error trap includedlambda.interp which was returning NaN under degenerate
circumstances.Surv objectpredict and coef with exact=TRUE. The
user is strongly encouraged to supply the original x and y
values, as well as any other data such as weights that were used in
the original fit.lognet when some weights are zero and x is sparsepredict.glmnet, predict.multnet and predict.coxnet,
when s= argument is used with a vector of values. It was not doing
the matrix multiply correctlyintercept optionglmnet.control for setting systems parameterscoxnetexact=TRUE option for prediction and coef functionsmgaussian family for multivariate responsegrouped option for multinomial familynewx and make dgCmatrix if sparselognet added a classnames component to the objectpredict.lognet(type="class") now returns a character vector/matrixpredict.glmnet : fixed bug with type="nonzero"glmnet: Now x can inherit from sparseMatrix rather than the very
specific dgCMatrix, and this will trigger sparse mode for glmnetglmnet.Rd (lambda.min) : changed value to 0.01 if nobs < nvars,
(lambda) added warnings to avoid single value,
(lambda.min): renamed it lambda.min.ratioglmnet (lambda.min) : changed value to 0.01 if nobs < nvars
(HessianExact) : changed the sense (it was wrong),
(lambda.min): renamed it lambda.min.ratio. This allows it to be
called lambda.min in a call thoughpredict.cv.glmnet (new function) : makes predictions directly from
the saved glmnet object on the cv objectcoef.cv.glmnet (new function) : as abovepredict.cv.glmnet.Rd : help functions for the abovecv.glmnet : insert drop(y) to avoid 1 column matrices; now include a glmnet.fit object for later predictionsnonzeroCoef : added a special case for a single variable in x; it was dying on thisdeviance.glmnet : includeddeviance.glmnet.Rd : includedglmnet_1.4.