evoFE 1.0.0

New Features

• Untouched confirmation holdout (holdout_frac): With evaluation_strategy = "cv", a stratified fraction of rows is excluded from the entire search and scored once after evolution. The result is reported as holdout_fitness and a new search_gap diagnostic (holdout minus validation score) on the evo_recipe object and in summary(), making search-level overfitting visible.
• Leakage-safe CV fold strategies (cv_strategy): New options "time" (contiguous chronological blocks ordered by time_col, so validation always lies in the future of training) and "group" (all rows of a group_col entity stay in one fold, balanced greedy assignment) alongside the default "random".
• Multi-fidelity evaluation (multi_fidelity, mf_sample_frac, mf_warmup_frac): Optional cost reduction in which warm-up generations screen individuals on row-subsampled folds; the most promising half is re-evaluated at full fidelity before any selection decision, keeping fitness comparisons apples-to-apples.
• New transformer date_diff: Signed difference in days between two datetime columns.

Breaking Changes

• The mbo_infill_opt = "ea" option is deprecated and ignored with a warning. The emoa package is no longer needed.

Enhancements

• Migrated Bayesian Optimisation backend from mlrMBO / ParamHelpers / smoof / lhs to mlr3mbo / paradox / bbotk. The public API (make_tunable(), lightgbm_mbo, xgboost_mbo) remains fully unchanged.
• Removed legacy packages mlr, randomForest, DiceKriging, and emoa from Suggested dependencies. The manual Kriging-to-RandomForest surrogate fallback has been replaced by mlr3mbo’s transparent internal surrogate management.
• Reduced package dependencies in Imports from 12 to 11.
• Added support for a Hybrid Island Population Model in evolve_features() via parameters islands, migration_interval, migration_rate, and gene_migration_prob.
  • Recipe-Level Migration: Exchanges complete successful feature recipes between islands in a Ring topology to preserve co-adapted interactions. Guarded against edge cases where dynamic population sizes shrink below migration_rate.
  • Gene-Level Migration (Injection): Periodically injects highly successful individual features (genes) from the ring neighbor’s best individual into the local mutation pool. Genes diffuse naturally through the ring over successive migrations, creating a gradient of innovation where each island builds hierarchically on its neighbor’s features.
  • Independent Local Stagnation: Dynamic population sizing and adaptive exploration rates are computed independently per island, allowing stagnant islands to expand and explore while active ones stay compact. Population growth is capped at 5x pop_size to prevent runaway expansion.
  • Verbose Reporting: Clear logging has been integrated to distinguish exactly which island and individual is being evaluated, including gene pool injection events.
• Added support for a Hybrid Feature Mask Approach for raw feature selection alongside evolved feature engineering:
  • Sigmoid Active Mask Initialization: Guides active feature masks of new individuals using baseline model importances centered at average feature importance.
  • Smart Mask Mutation Operators: Supports multi-column toggles scaled via a dynamic geometric distribution, single raw feature toggles with temperature-scaled importance probability, and complete mask recalculation.
  • With-Replacement Multivariate Sampling: Employs with-replacement sampling and deduplication for multivariate transformers to concentrate inputs around active, important features.
  • Pruning Safety Floor Restoration: Safeguards evaluations by automatically restoring raw features to meet the min_active floor if all genes are pruned during evaluation.
  • Cache Hit Logging Reduction: Silenced redundant (Cache Hit) prints for centroid distance components to reduce console verbosity.

evoFE 0.2.0

Bug Fixes

• Fixed CatBoost evaluator writing to disk by enforcing allow_writing_files = FALSE and redirecting any diagnostic files to tempdir().
• Fixed genotype corruption issue by treating failing or constant features as lethal mutations during training.
• Fixed -Inf fitness error by passing allow_prune to evaluate_fitness for pooled features.
• Fixed NA handling in quantile binning.
• Removed explicit [Cache Hit] printing from evaluate_pop.
• Fixed is_logits flag in xgboost custom metric evaluation and aligned implementation.
• Fixed xgb_feval: XGBoost custom_metric always receives raw logits, set is_logits=TRUE unconditionally.
• Fixed Laplace smoothing in TS-refinement formulation.
• Fixed mutation bug: the p (power transform) and q (groupby_quantile) parameters were missing from mutation logic and are now actively mutated.
• Fixed multi-component UMAP genes to sample shared parameters once per transformer addition, ensuring consistency across components.

Enhancements & Refactoring

• Optimized stateful clustering transformers: vectorized the fallback KNN distance search (3.8x speedup), integrated UMAP downsampling (7x speedup on fitting), and enabled UMAP prediction caching (1310x speedup on cache hits).
• Reduced memory footprint in cross-validation loops by removing redundant data.table::copy() calls.
• Hardened .cluster_prep_x() and .cluster_knn_apply() to safely handle edge-case configurations like vector-valued option inputs and NA distance matrices.
• Improved TS-refinement evaluation metric and taboo threshold calculations.
• Set XGBoost default min_child_weight=20 to match LightGBM min_data_in_leaf and prevent overconfident predictions.
• Removed seed parameter and set.seed() calls inside package code for CRAN compliance.
• Added input validation for split_ids: length check, label validation, and automatic evaluation_strategy switching to "split".
• Display actual split proportions (computed from split_ids) in the evolution header.
• Optimised .cluster_prep_x() deduplication using duplicated() instead of data.table grouping for improved performance.
• Clamped infinite values and imputed NA/NaN in TS-refinement log-likelihood calculations.
• Added globalVariables() declarations for data.table NSE symbols to eliminate R CMD check NOTEs.
• Integrated dynamic n_neighbors (Poisson mean 15) and dens_scale (uniform [0,1]) parameters into the UMAP transformer.
• Updated gene_to_state_formula to include all configuration parameters in the cache key to prevent collision.
• Replaced the flawed taboo threshold formula with a piecewise, hybrid scale-invariant approach for proper scaling across bounded and unbounded metrics.

Documentation

• Added CRAN status badge and updated installation instructions for CRAN release.
• Documented parameters for CRAN compliance.
• Reformatted Rd documentation to stay within 80-character line width.
• Improved split_ids documentation with usage examples.
• Added allowed_transformers parameter documentation to mutate() and initialize_population().
• Added missing @param datetime_cols documentation in create_individual() and initialize_population().
• Created central package-level documentation accessible via ?evoFE outlining all package options (evoFE.redundancy_cor_threshold, evoFE.importance_threshold, evoFE.max_clustering_size, evoFE.threads, and evoFE.verbose).
• Added \donttest{} @examples blocks to the internal-but-exported functions: create_individual(), mutate(), crossover(), and tournament_select().

evoFE 0.1.0

• Initial CRAN submission.