As of February 2024:
hal9001 and origami in DESCRIPTION to match the
latest CRAN releases, resolving bugs related to Matrix v1.6-2 as reported
at https://github.com/tlverse/hal9001/issues/109.As of May 2021:
pkgdown site in preparation for JOSS paper.As of October 2021:
LazyData field in DESCRIPTION since no data directory present.glmnet in documentation to avoid adding the package
to dependencies.As of October 2021:
smoothness_orders as a named argument to haldensify,
fit_haldensify, and cv_haldensify, with a default of zero. This was
previously passed to hal9001::fit_hal via ... arguments.As of September 2021:
hal9001::fit_hal() in keeping with updates
to that package, for compatibility with its v0.4.0 CRAN release.smoothness_orders argument of hal9001::fit_hal() previously was set
through the ... argument of haldensify; however, it has now been made a
named argument to both haldensify and the internal cv_haldensify and
fit_haldensify functions. The default is set to zero, for indicator basis
functions, which differs from the default of hal9001::fit_hal() as of its
v0.4.0 release.As of April 2021:
hal9001::fit_hal() in order to correctly use
the pared-down interface introduced in v0.4.0, contributed by @rachaelvp.A
has been altered to be multiples of sqrt(length(A)).As of April 2021:
haldensify arguments (removal of hal_max_degree as a named
argument) to simplify and better match use of fit_hal in hal9001 v0.3.0+.
This overhaul also included the addition of ... arguments, now passed
through haldensify and fit_haldensify to cv_haldensify, allowing all
internal calls to hal9001::fit_hal() to specify the same arguments be
passed for the fitting of HAL models.n_bins, now setting this to
(much) larger values that are themselves based on the sample size. This is in
accordance with evidence from simulation experiments indicating that higher
values of n_bins lead to significantly improved density estimates.trim and trim_dens to predict.haldensify to support
the use of truncation more transparently. While the default was to set
predictions for values of new_A outside the training support to zero, this
has been changed to avoid trimming and, when the choice is made to trim the
predictions, to set this value to 1/sqrt(length(new_A)).print.haldensify for a more user-friendly display
of the prediction procedure's output, including the selected number of bins,
the CV-selected choice of the regularization parameter, and the summary
of the fitted HAL model.As of February 2021:
plot.haldensify to simplify visualizing the empirical
risks of the sequence of HAL-based conditional density estimators across the
grid of the regularization parameter, and necessary changes to the vignette.type argument in the
plot.haldensify method. Not yet implemented.dplyr.haldensify() to full-data fit by explicitly passing n_folds = 1.n_bin and
grid_type are fixed; add related assertion check in predict() when
haldensify() skips cross-validation (since lambda selection skipped).haldensify() and predict() to clarify variable passing.predict() method to truncate small conditional density estimates to
a minimum value of [1 / sqrt(n)], based on the prediction set sample size.As of January 2021:
hal_basis_list to haldensify(), allowing for a HAL
basis produced by fit_hal() to be passed into the HAL regression used for
density estimation. This facilitates reduced computational overhead when
requiring external cross-validation of nuisance functions (e.g., CV-TMLE) as
well as working with bootstrap samples.hal_max_degree to haldensify(), allowing for control
of the highest degree of interactions considered in the HAL model for density
estimation. Like the above, this can reduce computational overhead.haldensify() by passing cv_folds to the n_folds
argument of fit_hal() when fitting HAL regression for density estimation.
Previously, cv_folds was only used in constructing cross-validation (CV)
folds for choosing tuning parameters, but the subsequent HAL regression was
fiex to use the default number of folds specified in fit_hal() to choose
the regularization parameter of the HAL regression for density estimation.
Now, both CV to choose density estimation tuning parameters and CV to choose
the lasso tuning parameter use the same number of folds.... to haldensify() so that arbitrary arguments can
be passed to fit_hal() for density estimation, when not already specified
as other arguments of the haldensify() constructor.use_future, specifying parallel evaluation
in a note instead."all" to the lambda_select argument of the predict()
method, allowing for predictions on the full (non-truncated) sequence of
lambdas fitted on to be returned.predict() method to 1/n instead of zero.As of January 2021:
W = NULL is set in haldensify().hal9001 dependency has been upgraded to v0.2.8 of that package, which
introduced breaking changes in the names of slots in fitted model objects.haldensify estimates, this
subsetting proved problematic for convergence of glmnet().predict() method's cv_select argument has been replaced in order to
better facilitate undersmoothing. The new argument lambda_select defaults
to the cross-validation selector but now easily allows access to the sequence
of undersmoothed density estimates (less restrictive regularization values).haldensify S3 output class have been changed
grid_type_tune_opt is now grid_type_cvselect,n_bins_tune_opt is now n_bins_cvselect, andcv_hal_fits_tune_opt is now cv_tuning_results.As of December 2020:
plan(transparent) has been changed to plan(sequential) based on
ongoing development in the future package ecosystem.As of June 2020:
As of May 2020:
haldensify for fitting HAL models has
been slightly abstracted and moved to the new function fit_haldensify.haldensify() wrapper function serves to cross-validate over choices of
the histogram binning strategy and the number of bins.haldensify() have been changed based on results of
simulation experiments.seed_int in haldensify() has been removed.cv_select, defaulting to TRUE, has been added to the
predict method, to make undersmoothing more accessible.