Scoring via random forests

Three different random forest models can be used to measure predictor importance.

Usage

score_imp_rf

score_imp_rf_conditional

score_imp_rf_oblique

Format

An object of class filtro::class_score_imp_rf (inherits from filtro::class_score, S7_object) of length 1.

Value

An S7 object. The primary property of interest is in results. This is a data frame of results that is populated by the fit() method and has columns:

name: The name of the score (e.g., imp_rf).
score: The estimates for each predictor.
outcome: The name of the outcome column.
predictor: The names of the predictor inputs.

These data are accessed using object@results (see examples below).

Details

These objects are used when either:

The predictors are numeric and the outcome is a factor/category, or
The predictors are factors and the outcome is numeric.

In either case, a random forest, conditional random forest, or oblique random forest (via ranger::ranger(), partykit::cforest(), or aorsf::orsf()) is created with the proper variable roles, and the feature importance scores are computed. Larger values are associated with more important predictors.

When a predictor's importance score is 0, partykit::cforest() may omit its name from the results. In cases like these, a score of 0 is assigned to the missing predictors.

Estimating the scores

In filtro, the score_* objects define a scoring method (e.g., data input requirements, package dependencies, etc). To compute the scores for a specific data set, the fit() method is used. The main arguments for these functions are:

object: A score class object (e.g., score_imp_rf).
formula: A standard R formula with a single outcome on the right-hand side and one or more predictors (or .) on the left-hand side. The data are processed via stats::model.frame()
data: A data frame containing the relevant columns defined by the formula.
...: Further arguments passed to or from other methods.
case_weights: A quantitative vector of case weights that is the same length as the number of rows in data. The default of NULL indicates that there are no case weights.

Missing values are removed by case-wise deletion.

In cases where the underlying computations fail, the scoring proceeds silently, and a missing value is given for the score.

Examples


library(dplyr)

# Random forests for classification task

cells_subset <- modeldata::cells |>
  # Use a small example for efficiency
  dplyr::select(
    class,
    angle_ch_1,
    area_ch_1,
    avg_inten_ch_1,
    avg_inten_ch_2,
    avg_inten_ch_3
  ) |>
  slice(1:50)

# Random forest
set.seed(42)
cells_imp_rf_res <- score_imp_rf |>
  fit(class ~ ., data = cells_subset)
cells_imp_rf_res@results
#> # A tibble: 5 × 4
#>   name       score outcome predictor     
#>   <chr>      <dbl> <chr>   <chr>         
#> 1 imp_rf -0.00283  class   angle_ch_1    
#> 2 imp_rf -0.00472  class   area_ch_1     
#> 3 imp_rf  0.0419   class   avg_inten_ch_1
#> 4 imp_rf  0.0604   class   avg_inten_ch_2
#> 5 imp_rf  0.000662 class   avg_inten_ch_3

# Conditional random forest
cells_imp_rf_conditional_res <- score_imp_rf_conditional |>
  fit(class ~ ., data = cells_subset, trees = 10)
cells_imp_rf_conditional_res@results
#> # A tibble: 5 × 4
#>   name                  score outcome predictor     
#>   <chr>                 <dbl> <chr>   <chr>         
#> 1 imp_rf_conditional -0.00889 class   angle_ch_1    
#> 2 imp_rf_conditional  0.0377  class   area_ch_1     
#> 3 imp_rf_conditional  0.199   class   avg_inten_ch_1
#> 4 imp_rf_conditional  0.616   class   avg_inten_ch_2
#> 5 imp_rf_conditional  0       class   avg_inten_ch_3

# Oblique random forest
cells_imp_rf_oblique_res <- score_imp_rf_oblique |>
  fit(class ~ ., data = cells_subset)
cells_imp_rf_oblique_res@results
#> # A tibble: 5 × 4
#>   name              score outcome predictor     
#>   <chr>             <dbl> <chr>   <chr>         
#> 1 imp_rf_oblique -0.0102  class   angle_ch_1    
#> 2 imp_rf_oblique  0.00764 class   area_ch_1     
#> 3 imp_rf_oblique  0.0901  class   avg_inten_ch_1
#> 4 imp_rf_oblique  0.0759  class   avg_inten_ch_2
#> 5 imp_rf_oblique -0.00673 class   avg_inten_ch_3

# ----------------------------------------------------------------------------

# Random forests for regression task

ames_subset <- modeldata::ames |>
  # Use a small example for efficiency
  dplyr::select(
    Sale_Price,
    MS_SubClass,
    MS_Zoning,
    Lot_Frontage,
    Lot_Area,
    Street
  ) |>
  slice(1:50)
ames_subset <- ames_subset |>
  dplyr::mutate(Sale_Price = log10(Sale_Price))

set.seed(42)
ames_imp_rf_regression_task_res <-
  score_imp_rf |>
  fit(Sale_Price ~ ., data = ames_subset)
ames_imp_rf_regression_task_res@results
#> # A tibble: 5 × 4
#>   name     score outcome    predictor   
#>   <chr>    <dbl> <chr>      <chr>       
#> 1 imp_rf 0.00246 Sale_Price MS_SubClass 
#> 2 imp_rf 0.00233 Sale_Price MS_Zoning   
#> 3 imp_rf 0.0115  Sale_Price Lot_Frontage
#> 4 imp_rf 0.00839 Sale_Price Lot_Area    
#> 5 imp_rf 0       Sale_Price Street