featR 0.2.0

A correctness release. Two adversarial reviews of every function and every documentation page found and fixed silent wrong answers, hangs, crashes, and inaccurate claims. Fixes that change behavior are listed first under each review. Where a result changed, the old result was wrong.

• featR now requires withr 2.4.0 or later, the first release that exports local_seed().
• Examples that run in under two seconds are no longer wrapped in \donttest{}, so R CMD check runs them.

First adversarial review

A full statistical, methodological, and programmatic review. Every finding was reproduced in R before it was fixed, and each fix has a regression test. This round also added the pkgdown site’s guides and articles.

Behavior changes

• fs_elastic()’s default alpha_seq is now seq(0.1, 1, by = 0.1). The old default included alpha = 0, pure ridge. Ridge never zeroes a coefficient, so whenever it won the tuning (it won on every seed tried with correlated predictors) every predictor was reported as selected. A user-supplied alpha = 0 that wins now warns.
• fs_lasso() gains lambda = c("min", "1se"). The default is unchanged; "1se" selects at lambda.1se for a sparser, more stable set, and details$lambda_used records which was used.
• summary() ranks scores by value rather than by absolute value, except for fs_lasso()’s signed coefficients. Negative permutation or Boruta importance (worse than noise) previously ranked above useful features. It also lists the kept side first when a threshold filter kept the low scores.
• fs_boruta() passes num.threads = 1 to Boruta’s importance engine, which otherwise used every core, contrary to featR’s sequential default.
• seed must be a whole number within the integer range. Fractional seeds were silently truncated (seed = 1.9 gave the same stream as seed = 1).
• Parallel paths restore the caller’s own foreach backend on exit, through one shared helper. Previously they forced the session back to sequential, or left a user-registered backend pointing at featR’s stopped cluster.

Statistical corrections

• fs_svm()’s SVM-RFE subset-size search was optimistically biased. It scored the top of a ranking built on all training rows, including each fold’s held-out rows, and scaled once on all rows. On pure noise it reported cross-validated accuracy above 90% (up to 100% with 200 noise predictors and 60 rows) where the truth is 50%. The elimination and the scaling are now re-run inside every fold. The best size’s score is still the maximum over the candidate sizes, so it remains mildly optimistic (about 60% on the same noise); use the held-out details$performance as the honest estimate.
• fs_bayes() no longer discards sampler and loo warnings. Convergence problems and high Pareto k values are collected and summarized after the search, saying whether the selected model is affected.
• fs_randomforest() pooled per-forest importance standard errors with randomForest::combine() when run in parallel. That inflated them by about sqrt(n_cores) and shrank scaled importance by the same factor. They are now pooled exactly across trees.
• fs_recursivefeature() over-reported held-out accuracy for character and logical targets. The test factor was rebuilt from the test rows alone, and caret::postResample() silently dropped predictions of classes absent from them.
• fs_supervised() computes the ANOVA F in closed form. Perfect separation scores Inf instead of a rounding-noise value near 1e31, only observed groups count toward the degrees of freedom, and the “essentially perfect fit” warning is gone. The constant-column check for correlation scores is relative to the data’s scale, so features measured in very small units are no longer scored NA.
• fs_infogain() clamps floating-point residue to zero, so exactly independent predictors score 0 and are not selected, and bins columns containing infinite values instead of failing.
• fs_pca() and fs_svd() use a relative zero-variance tolerance. Columns that are constant up to round-off are no longer scaled up to take a full share of a component: fs_pca() drops them with a warning, and fs_svd() rejects them with an error naming them when it is asked to scale. Columns in very small units are kept.
• fs_svd() falls back to the exact solver when RSpectra fails to converge.
• fs_mars() returns regression predictions and R2 as plain numbers rather than 1 x 1 matrices.

Bug fixes

• Non-syntactic column names now work. fs_mars() never selected such predictors, and fs_lasso(), fs_elastic(), and fs_stepwise() returned them backticked, so data[selected(res)] failed. fs_svm() carried the same backticks into its encoded feature names. fs_recursivefeature() crashed on them with caret::lmFuncs. Formula construction now escapes backslashes and backticks inside names.
• fs_chi() and fs_infogain() reject duplicated column names. The second copy was previously never tested.
• fs_lasso() and fs_elastic() give a clear error for a single design-matrix column instead of glmnet’s.
• fs_stepwise() accepts a user-supplied scope instead of erroring.
• fs_mars(search = "random") with a factor outcome no longer fails on the first call in a fresh session.
• fs_randomforest() rejects invalid sampsize entries instead of silently growing every tree on one row, imputes logical predictors, and warns when a numeric target is modeled with the default task = "classification".
• fs_recursivefeature() handles unused factor levels in the target, rejects fractional sizes and NA targets, reports NA test predictions, and requires randomForest when its defaults need it.
• fs_correlation() validates the type of diag_value. Its documentation now describes the pruning rule accurately: strongest pair first, not caret::findCorrelation()’s scan order.
• fs_pca() keeps logical and date columns as labels instead of dropping them silently.

Second adversarial review

A second review, then a sweep of every bug class it found across the whole package. Every finding was reproduced in R by one checker and confirmed by a second before it was fixed, and each fix has a regression test.

Behavior changes

• fs_randomforest(): control$strata must now be the name of a column. It is read from the training rows. A vector was silently misaligned with the training split. An NA in the strata column is an error.
• fs_recursivefeature(): scores is now each candidate’s full-size importance averaged over every resample. caret’s varImp(rfe) averaged only over the resamples where a variable survived to the optimal size, and left some candidates unscored. It remains in details$variable_importance.
• fs_recursivefeature(), fs_svm(select_method = "rf_rfe") and fs_randomforest(control = list(drop_zerovar = FALSE)) remove (or reject) predictors that are constant on the training rows. These hung randomForest indefinitely on classification targets. A predictor that is constant only inside one cross-validation resample (a rare indicator) is now set aside for that resample’s fits, with importance 0.
• When a dummy-variable name such as g1 (factor g, level 1) collides with another column or the target, fs_recursivefeature(), fs_svm(), fs_lasso(), fs_elastic() and fs_stepwise() now stop with an error. The real predictor was otherwise silently replaced by the dummy.
• Inf/-Inf in the target or predictors is now a clear error in fs_elastic(), fs_svm(), fs_recursivefeature(), fs_randomforest(), fs_boruta() and fs_bayes(), and fs_pca() now errors on it too. These previously failed inside the engine or were silently zeroed.
• User row indices that featR’s own NA drop or split would misapply are refused: index, indexOut and now indexFinal (fs_elastic(), fs_recursivefeature()).
• Factor levels absent from the training rows are dropped before encoding in fs_recursivefeature(), fs_svm() and fs_randomforest(), so they no longer become all-zero candidate features.
• fs_lasso() reports coefficients below 1e-10 times the largest standardized coefficient as 0, and gives numerically constant columns a coefficient of 0. Round-off residue was being selected.
• fs_stepwise() scores are keyed by term, like selected. A factor term scores the largest |t| among its coefficients. Per-coefficient values are in details$coefficient_scores.
• fs_mars() and fs_randomforest() return NA for ROC AUC when the held-out rows contain one class (fs_randomforest() also warns). They used to fail after training.
• fs_svm(select_method = "rf_rfe", n_features = k) keeps the top k by the reported scores. Its fallback screens on permutation importance, not impurity importance, which almost every predictor passed.
• fs_bayes(seed = s) now seeds the MCMC and no longer advances the caller’s RNG. Failed loo() fits and loo_compare() failures are now warnings. details$n_loo_failed and details$n_compared are added.
• fs_chi(parallel = TRUE) respects a future plan you already set, and restores the previous plan even when worker start-up fails.
• summary() lists selected features first for fs_mars() and fs_correlation(prune = TRUE).
• Stricter, earlier validation:
  • rfe_control/train_control methods, number >= 2, and complete functions sets.
  • Full trainControl() lists in fs_elastic().
  • p_adjust_method in fs_chi().
  • control$positive_class in fs_randomforest().
  • Engine requirements such as e1071, checked before any fitting.
• fs_elastic() reports the metric caret actually optimized when it fell back from an unknown metric.

Statistical corrections

• fs_elastic() selected predictors whose coefficient is zero at the tuned lambda. glmnet coefficients were interpolated between path knots; they are now computed exactly.
• fs_mars() dropped predictors that are in the final earth model when their GCV importance was <= 0. Selection now reads the model’s terms.
• fs_svm(select_method = "rf_rfe") weighted resamples by survivorship when averaging importance. It now averages each predictor’s full-size importance once per resample.
• Seeded parallel fs_recursivefeature() and fs_svm() runs are now reproducible. caret’s per-resample seeds made two such runs differ.
• fs_chi() warns when its Monte Carlo p-value floor, 1/(simulation_B + 1), means a feature cannot reach sig_level after multiplicity adjustment.
• Constant-column tests across fs_elastic(), fs_svm(), fs_lasso(), fs_infogain(), fs_pca(), fs_svd() and fs_supervised() are relative to the data’s scale. Round-off-constant columns are no longer scaled to full weight, and columns in extreme units are no longer mistaken for constants.
• fs_randomforest() regression R2 is NA for a held-out target that is constant up to round-off. It was about -1e29. Its kappa is computed directly and is no longer silently NA without e1071. A factor target in regression uses its labels, not its level codes.
• Optimism is documented where a reported quantity reuses the rows that selection saw:
  • fs_randomforest() details$oob with a supervised feature_select.
  • fs_svm() model$results and the chosen size’s size_scores.
  • fs_bayes() details$best_elpd.
  • The best-of-grid fs_elastic() metric.
  The held-out details metrics are the honest estimates. Metric definitions (for example caret’s Rsquared is a squared correlation) are stated.

Bug fixes

• Names:
  • Duplicated, empty or NA column names are rejected rather than silently dropping or merging a column, across all the wrappers, fs_boruta(), fs_pca(), fs_chi(), fs_infogain() and fs_supervised().
  • fs_recursivefeature(handle_categorical = TRUE) and fs_mars() classification returned backticked names, or dropped non-syntactic predictors.
  • Fixed-name internal columns (.mars_train_idx, desired_total, residuals, fitted_values, iso_week_id, a predictor named data) no longer clash with user columns.
  • Matrix input keeps its own column names.
• fs_mars() classification with n_cores > 1 failed on the workers. Random search left earth’s dependencies attached.
• Clear featR errors replace cryptic engine errors for:
  • one-level factor predictors;
  • ordered, single-class, empty-label and unused-level targets;
  • saturated stepwise models (backward/both; a forward search that ends on a perfect fit warns);
  • one-row splits;
  • list columns;
  • every predictor removed by filtering;
  • factor predictors with caret::lmFuncs.
• Restoring the foreach backend no longer marks a never-registered session as registered.
• fs_lasso()’s lambda argument moved before the housekeeping arguments.

featR 0.1.0

First release. Unifies the feature_selection script collection into a package with 16 exported fs_*() functions (plus the selected() accessor) sharing one calling convention and one return type.

Calling convention

All selection functions take the form:

fs_<method>(data, target, <method options>, ...)

data is always first and target (a single column-name string) always second, for the functions that have an outcome. Housekeeping arguments come last: every function takes verbose, functions that use randomness take seed = NULL, and functions that can run in parallel take n_cores and/or parallel. 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).

Return value

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.

Statistical corrections

• 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.
• Class upsampling in 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.

Bug fixes

• 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 mislabeled 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.
• In fs_svm(select_method = "rf_rfe"), when random-forest RFE fails, the fallback keeps every predictor with positive impurity importance, and only falls back to the top n_features (1 when n_features = NULL) 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.

Package conventions

• Modeling engines are Suggests; each function checks for what it needs.
• Functions never seed the RNG unless seed is supplied, and restore the previous RNG state afterward.
• Execution is sequential by default; worker counts are opt-in and capped.
• No library(), install.packages(), .GlobalEnv writes, or log files in package code.