Regression ensembles without mlr3pipelines

Version 1.2.0 replaces the graph and preprocessing operators formerly supplied by mlr3pipelines with internal code. The core runtime retains mlr3 and mlr3learners for regression interfaces, mlr3tuning, paradox and bbotk for tuning, data.table for task backends, and R6 for the internal learner. No mlr3pipelines installation is required, including for saved new models.

The six supported learner families still require their statistical backends: glmnet, ranger, xgboost, rpart, kknn and e1071 (for svm). Only install those you request. Reimplementing those modelling algorithms is outside the scope of the preprocessing replacement. The default algorithm vector requests all six.

Correlation and variance filtering use internal code. The historical default information_gain uses optional mlr3filters and FSelectorRcpp; relief also requires FSelectorRcpp, carscore requires care, and cmim requires praznik. find_correlation uses mlr3filters. Mixed-model encoding (encodelmer) uses lme4 only when requested. Missing optional dependencies produce an installation message identifying the selected feature. knitr builds this vignette and is not needed for ordinary fitting. Tests use base R without a test framework.

A runnable example

This example uses two inexpensive backends and internal variance filtering. Install rpart and e1071 if needed. Vignette builds without these optional packages show the code and skip the model fit.

set.seed(12)
training <- data.frame(
  size = rnorm(60), age = runif(60, 0, 20),
  group = rep(c("north", "south", "west"), 20),
  active = rep(c(TRUE, FALSE), 30)
)
training$value <- 4 * training$size - training$age / 3 + rnorm(60)
training$size[c(3, 10)] <- NA_real_

result <- sense(
  training, "value", algos = c("rpart", "svm"),
  selected_filter = "variance", selected_n_feats = 4,
  missing_fusion = TRUE, ratio = 0.7,
  budget = 1, n_evals = 1, seed = 42
)

The one-trial budget keeps the example fast; increase it for a substantive search. Trials are evaluated one at a time. Random search respects both budget and n_evals; grid search respects n_evals and resolution. Stagnation and time limits are checked between trials, so a running fit is not interrupted.

result$test_metrics

#>       mse      rmse       mae      mape      mdae       rae       rse      rrse 
#> 99.612872  9.980625  8.042884  8.131120  7.204282  3.184250 13.693000  3.700405 
#>     smape 
#>  1.352413

result$selected_n_feats

#> $feats
#> [1] "x2_1" "x4_1" "x1_1" "x1_2"
#> 
#> $score
#>       x1_1       x1_2       x2_1       x3_1       x3_2       x3_3       x4_1 
#> 0.25423729 0.25423729 1.00000000 0.22598870 0.22598870 0.22598870 1.00000000 
#>   missing4 
#> 0.03276836

new_data <- training[1:3, setdiff(names(training), "value")]
new_data$group[1] <- "unseen-region"
result$model_predict(new_data)

#>   value
#> 1 -5.80
#> 2 -2.48
#> 3 -3.84

result$time_log

#> Time difference of 2.148895 secs

What is learned, and where

Outer resampling assesses the entire tuning and fitting procedure. Each inner tuning split trains an independent internal ensemble. Within each ensemble, three-fold cross-validation produces a prediction for every training row from base learners that did not train on that row. The meta learner trains on these predictions; super = "avg" takes their arithmetic mean. A numerical benchmarking argument selects that many base algorithms using these out-of-fold scores, respecting whether the metric is minimized or maximized. Selection is repeated inside each training split.

Every stacking fold fits its own character-level collapse, imputation distributions, scaling, factor encodings and feature selection. In particular, target-based encodings and supervised filters cannot see the held-out outcomes. For deployment the bases are refitted on their complete training data, and the prediction function uses the final ensemble trained on all supplied rows. sampling_rate subsamples base-learner training rows after preprocessing; validation and prediction rows are not subsampled.

Numeric imputation draws from observed training values (sample) or their histogram (hist). A completely missing numeric column is imputed to zero and removed if constant. scale uses the training mean and standard deviation; range maps the training range to zero--one; nop leaves numeric values unscaled. Character levels beyond collapse_char_to are grouped using training frequencies, with one slot reserved for the collapsed group. Logical columns are categorical. A separate collision-safe level represents missing values.

Factor encodings include one-hot and treatment, polynomial, sum and Helmert contrasts; impact encoding uses smoothed differences from the training target mean. encodelmer fits a random-intercept model using lme4. New categories receive zero contrasts or zero impact, or the global mixed-model intercept. Prediction reuses fitted states and column order; it never refits them. Missing required predictor columns produce an error. A data set with no nonconstant encoded predictors cannot be fitted.

Migration and reproducibility

The public function arguments and nine named result elements are retained. The plot element is now a lightweight sense_pipeline object instead of an interactive HTML widget; call plot on it to draw the architecture. time_log is a difftime in seconds, replacing a lubridate period. java_mem is accepted for compatibility but does not change Java options. Internal model layouts have changed, so code inspecting GraphLearner or PipeOp state must migrate. Old serialized graph models still need their original dependencies: refit them with the new version before retiring mlr3pipelines.

benchmark_error now reports final-fit out-of-fold scores rather than an up-front holdout benchmark. selected_n_feats reports encoded names and filter scores from the first final base model; independently fitted bases can select different features after stochastic imputation. These states are stored with each base model. Changed fold boundaries and preprocessing mean predictions are not expected to be numerically identical to version 1.1.0.

The nine test metrics retain their formulas: squared, absolute, percentage, median and relative errors. They use the rounded testing predictions; model_error uses mlr3's unrounded predictions. Zero denominators retain ordinary R semantics (Inf or NaN), including constant targets for relative errors and zero truth for percentage errors.

Fitting with the same seed is reproducible and restores the caller's random state. Stochastic imputation during subsequent prediction uses the caller's random state: call set.seed before predicting if new data contain missing numbers and identical repeated predictions are required. saveRDS and readRDS preserve new fitted results, including their prediction functions.

Regression ensembles without mlr3pipelines

A smaller runtime

A runnable example

What is learned, and where

Migration and reproducibility