First release. Unifies the feature_selection script
collection into a package with 16 exported functions sharing one calling
convention and one return type.
All selection functions take the form:
fs_<method>(data, target, ..., seed = NULL, verbose = FALSE, n_cores = 1L)data is always first and target (a single
column-name string) always second. This replaces the previous mix of
response_col, target_var,
target_col, responseName,
response_var, dependent_var, and
x/y argument pairs. Other renames:
p -> train_ratio,
predictor_cols -> predictors,
out -> output,
log_progress/show_progress/doTrace
-> verbose,
cores/temp_multisession ->
n_cores, and control$split_ratio ->
control$train_ratio in fs_randomforest().
fs_randomforest() also moves seed and
n_cores out of its control list into real
arguments. Removed arguments that never worked or had no effect:
early_stop_threshold (fs_bayes()),
early_stop and feature_funcs
(fs_recursivefeature(), the latter now passed as
rfe_control$functions), seed and
return_models (fs_stepwise()),
memoise_result (fs_svd()),
auto_install (fs_randomforest()), and
method (fs_mars(), which is earth-only).
The 14 selection functions return an fs_result object
with selected, scores, method,
task, model, details, and
call, plus print(), summary(),
and selected() methods. Method-specific output moved into
details. fs_pca() and fs_svd()
are dimensionality reduction and keep their own decomposition
structure.
fs_svm() gained a real SVM-RFE implementation
(linear-kernel weight-vector ranking with a refit at each elimination
step, plus cross-validated size selection). The previous behavior —
random-forest RFE — is still available via
select_method = "rf_rfe".fs_correlation() gained prune (default
TRUE): selected is now the reduced
non-redundant set rather than both members of every correlated
pair.fs_boruta() prunes correlated features by Boruta
importance, keeping the stronger member of a correlated group instead of
deferring to a blind correlation heuristic.fs_bayes() selects with loo::loo_compare()
and a 1-SE parsimony rule (rule = "1se", default) instead
of the raw elpd maximum. The rule accepts either container
loo_compare() may return (matrix or data.frame); an earlier
is.matrix() guard would have silently disabled it under any
loo release that returns a data.frame.fs_infogain() gained
normalize = "gain_ratio" to correct information gain’s bias
toward high-cardinality predictors, and now discretizes the target once
so scores are comparable across features.fs_lasso() no longer mean-imputes by default
(impute = "none"); full-data imputation leaked across
cross-validation folds. Scores are now standardized coefficients.fs_elastic() fits PCA inside each resample via caret’s
preProcess instead of once on the full data.fs_randomforest() runs the
control$feature_select hook on the training split
only.fs_mars() and fs_svm()
happens within resampling folds rather than before
cross-validation.fs_recursivefeature() evaluates the selected subset on
a held-out test split and trains its final model on training rows
only.fs_bayes() reads the model labels
loo::loo_compare() produces from its model
column when present, falling back to row names. loo 2.10.0 moved those
labels out of the row names, which would otherwise have made the
selection rule map comparison rows to the wrong candidate
models.
fs_recursivefeature() coerces character and logical
predictors to factors before fitting the one-hot encoder, so
the encoder is not silently refitted on the test rows
(caret::dummyVars() records levels only for columns that
are already factors).
fs_supervised() honors na_rm = FALSE on
the ANOVA path, where stats::lm()’s own NA handling
previously made it behave like na_rm = TRUE.
fs_unsupervised() returns an undefined score rather
than an error for method = "iqr" with
na_rm = FALSE.
fs_lasso() computes standardized scores by position
rather than by column name, so a design matrix with duplicated names
cannot pair a coefficient with the wrong standard deviation. Its
nfolds minimum is now 3, matching glmnet.
fs_mars() checks for MLmetrics before selecting
caret’s multi-class summary, instead of a test that could never fail;
multi-class targets no longer fail after the resamples have been
computed.
fs_correlation() rejects duplicated column names
instead of silently resolving every lookup to the first match, which
could replace one column’s correlations with another’s and drop both
members of the pair.
fs_boruta() validates maxRuns against
Boruta’s own minimum of 11, so the error names the featR argument rather
than surfacing from the dependency.
fs_lasso() reports non-finite predictor values as a
user-input problem, naming the offending columns, rather than as an
internal error.
fs_elastic() no longer errors with “missing value
where TRUE/FALSE needed” when a tuning result column contains
NA.
summary() on an fs_result ranks p-value
scores ascending, so fs_chi() lists its most significant
features first instead of last.
print() on an fs_result reports the
recorded candidate count in preference to the number of scored features,
so methods that score only a subset (such as
fs_recursivefeature()) no longer report “Selected 2 of
2”.
The fs_result constructor rejects an unnamed numeric
scores vector and a zero-length task, both of
which print() and summary() could not
display.
fs_bayes() samples predictor combinations without
enumerating them, so sample_combinations now works at the
scale it exists for: previously every subset was materialized first,
which exhausted memory past roughly 25 predictors. An unbounded search
over a very large subset space now errors with instructions instead of
dying on allocation.
fs_bayes() validates brm_family, so
passing a family generator without parentheses reports the mistake
instead of “object of type ‘closure’ is not subsettable”.
fs_bayes() supports families whose
fitted() is a 3-D array (categorical, multinomial,
multivariate). Selection proceeds; the in-sample MAE and RMSE, which are
undefined for those responses, are reported as NA rather
than causing an error that was previously mislabelled as a sampling
failure.
fs_svm() errors rather than ranking features from a
partially recovered weight vector, and reports a degenerate
elimination-step fit as a featR error naming the likely cause instead of
surfacing kernlab’s.
fs_recursivefeature() sets reproducible RNG streams
on its parallel workers, so two seeded parallel runs agree, and stops
its cluster if backend registration fails.
fs_pca() rejects scale_data = TRUE with
center_data = FALSE, which the two engines handled
differently, and validates num_pc against the large-data
engine’s own limit.
fs_randomforest() replaces only the NA
entries of a per-class control$sampsize, instead of
discarding the sizes that were supplied.
fs_infogain() refuses to expand a date column when
the resulting
<col>_year/_month/_day name
already exists. Previously that column was overwritten in place,
silently, including when it was the target.
fs_infogain() caps the automatic bin count at the
number of observations. A near-constant column beside one extreme
outlier drove the Freedman-Diaconis count past the integer range, which
became NA and made cut() fail with “invalid
number of intervals”.
fs_chi() reports correction_applied
from what stats::chisq.test() actually did. A 2x2 table
sitting exactly at expectation receives no Yates correction even when
one is requested, and the row previously claimed otherwise.
fs_correlation() registers its cluster teardown
before registering the parallel backend, so a failure there cannot leak
the cluster, and restores the caller’s own foreach backend instead of
forcing sequential execution.
fs_elastic() sets allowParallel from
whether featR created a cluster. caret’s default of TRUE
meant a single-worker call would dispatch resamples to whatever backend
the caller had registered elsewhere.
fs_svm()’s min_keep is a floor rather
than a quota: when random-forest RFE fails, the fallback keeps every
predictor with positive impurity importance and only drops to
min_keep top-ranked predictors if fewer qualify. It
previously returned exactly one predictor.
fs_mars() no longer fails on every call (a
data.table was indexed with the matrix returned by
caret::createDataPartition()).
fs_mars() sanitizes factor levels with
make.names(unique = TRUE), so distinct classes such as
"class 1" and "class.1" can no longer
merge.
fs_lasso() accepts data frames containing
NA (the previous model.matrix() call silently
dropped those rows and then aborted).
fs_pca() computes variance explained against total
variance in the large-data path, labels its loadings, and no longer
overflows on very large inputs.
fs_correlation() reports point-biserial correlations
with the correct sign.
fs_randomforest() stratifies on the user’s target
rather than any column literally named "target", imputes
test-only missing values, and accepts character predictors.
fs_svd() errors on invalid arguments instead of
silently repairing them.
seed is supplied,
and restore the previous RNG state afterwards.library(), install.packages(),
.GlobalEnv writes, or log files in package code.