Package {sense}

sense

Description

Stacked ensembles for regression using mlr3 and internal preprocessing.

Usage

sense(
  df,
  target_feat,
  benchmarking = "all",
  super = "avg",
  algos = c("glmnet", "ranger", "xgboost", "rpart", "kknn", "svm"),
  sampling_rate = 1,
  metric = "mae",
  collapse_char_to = 10,
  num_preproc = "scale",
  fct_preproc = "one-hot",
  impute_num = "sample",
  missing_fusion = FALSE,
  inner = "holdout",
  outer = "holdout",
  folds = 3,
  repeats = 3,
  ratio = 0.5,
  selected_filter = "information_gain",
  selected_n_feats = NULL,
  tuning = "random_search",
  budget = 30,
  resolution = 5,
  n_evals = 30,
  minute_time = 10,
  patience = 0.3,
  min_improve = 0.01,
  java_mem = 64,
  decimals = 2,
  seed = 42
)

Arguments

Details

Preprocessing and three-fold stacking are implemented internally without mlr3pipelines. Each stacking fold learns its own preprocessing state, including supervised encodings and filtering. Benchmark selection is repeated inside each training split. Outer resampling evaluates the complete tuning procedure. Numeric imputation samples from observed training values or their histogram; an entirely missing numeric column uses zero and is removed if constant. Unseen categories use a zero contrast or impact, or the fitted intercept for encodelmer. Character levels are collapsed using training frequencies only. Correlation and variance filters are internal. Other filters require the optional mlr3filters package and its method-specific backend. Learner backends and lme4 (only for encodelmer) must be installed when requested. Percentage and relative errors retain ordinary R division semantics: zero denominators can produce Inf or NaN. test_metrics uses rounded predictions; model_error uses unrounded mlr3 predictions. Random state is restored after fitting. Imputation at prediction time samples using the caller's random state.

Value

This function returns a list including:

• benchmark_error: out-of-fold comparison of base learners from the final fit, or NULL when benchmarking is "all".

• resampled_model: mlr3 standard description of the analytic pipeline.

• plot: a sense_pipeline object; use plot(result$plot) for a base R diagram.

• selected_n_feats: selected encoded feature names and scores from the first final base model; each base model has its own fitted preprocessing state.

• model_error: error measure for outer cycle of cross-validation.

• testing_frame: data set used for calculating the test metrics.

• test_metrics: metrics reported are mse, rmse, mae, mape, mdae, rae, rse, rrse, smape.

• model_predict: prediction function using the ensemble refitted on all supplied data.

• time_log: elapsed computation time as a difftime in seconds.

Author(s)

Giancarlo Vercellino giancarlo.vercellino@gmail.com

Maintainer: Giancarlo Vercellino giancarlo.vercellino@gmail.com

See Also

Useful links:

• https://mlr3.mlr-org.com/

Examples

## Not run: 
sense(benchmark, "y", algos = c("glmnet", "rpart"))

## End(Not run)

benchmark data set

Description

A data frame for regression task generated with mlbench friedman1.

Usage

benchmark

Format

A data frame with 11 columns and 150 rows.

Source

mlbench, friedman1

Plot the sense preprocessing and stacking architecture

Description

Plot the sense preprocessing and stacking architecture

Usage

## S3 method for class 'sense_pipeline'
plot(x, ...)

Arguments

Value

The pipeline, invisibly.