featR

A unified R package of feature-selection methods: statistical filters, regularization, model-based wrappers, and dimensionality reduction — all behind one calling convention and one return type.

Heavy modeling engines (brms, caret, glmnet, randomForest, ...) are optional Suggests; each function checks for what it needs and tells you what to install.

Documentation: https://elkronos.github.io/featR/. It has the function reference, a getting-started guide, method guides for every family, and articles on validation and reproducibility.

Installation

# install.packages("devtools")
devtools::install_github("elkronos/featR")

The convention

Every selection function takes the data first and the name of the target column second, and returns the same object:

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

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. The method guide has a table showing exactly which function accepts which. Functions with no outcome, fs_unsupervised(), fs_correlation(), fs_pca(), and fs_svd(), take no target.

library(featR)

res <- fs_lasso(mtcars, "mpg", nfolds = 5, seed = 1)

res               # <fs_result> lasso (regression), with the selected features
selected(res)     # character vector of chosen features
res$scores        # per-feature scores, comparable within a method
res$model         # the fitted model (fs_lasso keeps it only with return_model = TRUE)
res$details       # everything method-specific
summary(res)      # ranked score table, selected features marked

Because the shape is shared, methods are interchangeable:

runs <- list(
  lasso    = fs_lasso(mtcars, "mpg", nfolds = 5, seed = 1),
  stepwise = fs_stepwise(mtcars, "mpg"),
  rfe      = fs_recursivefeature(mtcars, "mpg", sizes = c(2, 4), seed = 1)
)
lapply(runs, selected)

Functions are sequential by default (parallelism is opt-in and capped at the detected core count), never seed the RNG unless you pass seed (and then restore your RNG state on exit), and signal progress with suppressible message()s.

Functions

Filters

Function	What it does
fs_chi()	Chi-square tests of association between categorical features and a categorical target, with p-value adjustment and Monte-Carlo fallback for sparse tables.
fs_correlation()	Flags variable pairs above a correlation threshold. prune = TRUE (default) returns the reduced non-redundant set; prune = FALSE returns every flagged pair member.
fs_infogain()	Information gain w.r.t. a target, for numeric (binned), categorical, and date features. normalize = "gain_ratio" corrects the bias toward many-leveled predictors.
fs_supervised()	Threshold filter: absolute correlation (numeric target) or ANOVA F (factor target).
fs_unsupervised()	Target-free threshold filter: variance, MAD, IQR, range, missing proportion, distinct-value count.

Regularization

Function	What it does
fs_lasso()	Cross-validated LASSO (glmnet). Scores are standardized coefficients; raw-scale ones are in details.
fs_elastic()	Elastic net over an alpha grid via caret + glmnet, with optional PCA fit inside each resample.

Wrappers and model-based

Function	What it does
fs_bayes()	Bayesian model comparison over predictor subsets (brms), ranked by LOO with a 1-SE parsimony rule.
fs_boruta()	Boruta all-relevant selection, with importance-aware correlation pruning.
fs_randomforest()	Random-forest permutation importance with held-out evaluation.
fs_recursivefeature()	caret RFE on a training split, honestly evaluated on held-out rows.
fs_stepwise()	Stepwise linear regression via MASS::stepAIC.
fs_svm()	SVM pipeline with true SVM-RFE (linear-kernel weight ranking, refit per elimination step); random-forest screening available via select_method = "rf_rfe".
fs_mars()	MARS (earth) with tuning and optional ROC/PR AUC.

Dimensionality reduction

These reduce dimensions rather than select features, so they return their own decomposition structure rather than an fs_result.

Function	What it does
fs_pca()	PCA (prcomp, or bigstatsr for large data) with loadings, scores, variance explained, optional plot.
fs_svd()	Exact or approximate (RSpectra) truncated SVD. Pass n_singular_values < min(dim(x)) to enable the approximate solver.

Notes on honesty

Several methods can silently flatter themselves. featR tries not to:

• Class upsampling happens inside resampling folds (fs_mars(), fs_svm()), never before.
• fs_recursivefeature() evaluates on rows the selection never saw, and its final model trains on training rows only.
• fs_elastic() fits PCA per fold; fs_lasso() refuses to mean-impute by default (impute = "none") because full-data imputation leaks across folds.
• fs_randomforest()'s feature_select hook runs on the training split only.
• Post-selection statistics (fs_stepwise() p-values) are labeled as invalid for inference rather than presented as if they were not.

License

MIT (c) Justin Chase. See LICENSE.md.