missForest NEWS

missForest 1.6.1 (2025-10-22)

Fixes and robustness improvements

• Fixed a parallelization edge case in parallelize = "variables" mode where empty index chunks could lead to uninitialized results.

• Added defensive guards (tryCatch, zero-length handling) to prevent “object ‘results’ not found” errors during variable-level parallel imputation.

• Improved handling of edge cases with few or single predictors (zero-covariate fallback).

• Introduced automatic worker capping to prevent errors when the number of parallel cores exceeds the number of variables.

• Enhanced internal stability for both ranger and randomForest backends in low-dimensional or degenerate datasets.

• Expanded the internal test suite with ADAPTS-like stress tests covering:

  • mixed-type small datasets
  • single-column and zero-covariate cases
  • both parallelization modes and backends

1.6 — 2025-10-13

Added

• Ranger backend (default): missForest() can now fit forests via ranger with backend = "ranger" (default) or retain legacy behavior with backend = "randomForest".
• Backend selector: New argument backend = c("ranger", "randomForest").
• Threading control: New argument num.threads (used by ranger). In parallelize = "variables" mode, per-variable ranger calls use num.threads = 1 to avoid oversubscription.

Changed

• Argument mapping for ranger:
  • ntree → num.trees
  • nodesize = c(num, fac) → min.bucket (regression/classification respectively)
  • sampsize (counts) → sample.fraction (overall or per-class fractions)
  • classwt → class.weights
  • cutoff emulated via probability forests + post-thresholding
  • maxnodes has no exact equivalent and is ignored by ranger (consider max.depth at the ranger level if needed).
• Defaults for node size: nodesize = c(5, 1) interpreted consistently across backends (numeric first, factor second).
• Parallelization behavior clarified:
  • parallelize = "variables": builds per-variable forests in parallel via foreach; ranger runs with num.threads = 1 inside each task.
  • parallelize = "forests": uses ranger’s internal threading (via num.threads) or combines sub-forests for randomForest.

Fixed

• Prevented name conflicts and S3 confusion by narrowing imports (e.g., stats::predict) and explicitly importing only what’s used.
• Resolved NOTE about undefined xntree in foreach by localizing/binding the chunk variable in the loop.
• Corrected usage of iterator utilities by importing itertools::isplitVector and iterators::idiv.
• Removed unused macro imports and unicode arrows in Rd files that triggered warnings.

Documentation

• Help pages refreshed with backend details, parameter mappings, and clear parallelization guidance.
• Cross-references now use package-qualified links:
  • \link[randomForest]{randomForest} and \link[ranger]{ranger}.
• Examples modernized (showing both backends and optional parallel usage).
• CITATION modernized to bibentry() with person()/c() and DOI.
• README rewritten to clearly explain capabilities, backends, and typical usage patterns.

Internal / Packaging

• DESCRIPTION
  • Version: 1.6 and current Date.
  • Added Authors@R.
  • Updated Imports to include only used packages (e.g., ranger, randomForest, foreach, iterators, itertools, doRNG, stats).
• NAMESPACE
  • Explicit, minimal imports to avoid clashes and CRAN checks.
  • Exported only user-facing functions: missForest, mixError, nrmse, prodNA, varClass.

Compatibility Notes

• Results across backends should be similar but not identical due to implementation differences (sampling semantics, splits, probability handling).
• maxnodes is ignored with backend = "ranger"; consider max.depth at the ranger level if tree depth control is required.