Reproducibility, randomness, and parallelism • featR

featR does not touch your random stream

No featR function calls set.seed() unless you pass seed. When you do, the seed applies for that call only. Your previous RNG state is restored when the function returns, even when it errors. Two consequences follow:

Two calls with the same seed give the same answer.
Calling featR with a seed does not change the random numbers your own code draws afterwards.

d <- data.frame(a = sin(1:50), b = cos(1:50), c = sin(1:50) + 0.1 * cos(3:52))

set.seed(2024)
expected <- runif(3)

set.seed(2024)
res <- fs_correlation(d, threshold = 0.9, sample_frac = 0.5, seed = 1)
observed <- runif(3)

# The seeded call left the caller's stream exactly where it was
identical(expected, observed)
#> [1] TRUE

Without a seed, randomized methods draw from your global stream like any other R function, so set.seed() before the call also works.

Which functions use randomness

Train/test splits: fs_randomforest(), fs_mars(), fs_recursivefeature(), fs_svm().
Cross-validation folds: fs_lasso(), fs_elastic(), fs_mars(), fs_recursivefeature(), fs_svm().
Model fitting: fs_randomforest() and fs_boruta() grow random forests.
Monte-Carlo p-values: fs_chi(), when expected cell counts are small.
Row sampling: fs_correlation(sample_frac < 1).
Subset sampling: fs_bayes(sample_combinations = ...).
Deterministic: fs_supervised(), fs_unsupervised(), fs_infogain(), fs_stepwise(), fs_pca(), and fs_svd() with the exact solver.

fs_bayes() is special. Its seed controls only which subsets are sampled. To make the MCMC reproducible, pass brm_args = list(seed = ...) as well.

Parallelism is opt-in and bounded

Every function runs on one core unless you ask for more. The rules are:

Worker requests are capped at parallel::detectCores(). Asking for 64 on a 4-core laptop gives 4.
Clusters featR creates are stopped when the call returns, including on error.
Parallel backends you registered yourself are restored afterwards.
With a seed, parallel workers get reproducible L’Ecuyer-CMRG streams. Two seeded parallel runs therefore agree with each other. They need not agree with a seeded sequential run, because the workers draw from different streams.

fs_randomforest(big_data, "y", task = "regression", n_cores = 4, seed = 1)
fs_lasso(big_data, "y", parallel = TRUE, n_cores = 4, seed = 1)

One caveat: when fs_randomforest() grows its forest on more than one worker, the pieces are merged with randomForest::combine(). That drops the out-of-bag error estimates, so details$oob is NULL. Use n_cores = 1 if you need the OOB numbers.

Optional dependencies

The only hard dependencies are data.table, parallel, stats, utils, and withr. Every modeling engine is in Suggests and is loaded only on the code path that needs it. A missing package gives an actionable error:

Error: Packages 'glmnet', 'Matrix' are required for lasso feature selection
but not installed. Install with: install.packages(c("glmnet", "Matrix"))

To install everything at once:

install.packages(c(
  "Boruta", "brms", "caret", "doParallel", "e1071", "earth", "foreach",
  "furrr", "future", "ggplot2", "glmnet", "kernlab", "loo", "MASS", "Matrix",
  "MLmetrics", "pbapply", "polycor", "pROC", "PRROC", "randomForest",
  "RSpectra", "bigstatsr"
))

fs_bayes() also needs a working C++ toolchain for Stan. See the brms installation notes.