Bagging functions

General suite of bagging functions for several models.

Usage

bagger(x, ...)

# Default S3 method
bagger(x, ...)

# S3 method for class 'data.frame'
bagger(
  x,
  y,
  weights = NULL,
  base_model = "CART",
  times = 11L,
  control = control_bag(),
  cost = NULL,
  ...
)

# S3 method for class 'matrix'
bagger(
  x,
  y,
  weights = NULL,
  base_model = "CART",
  times = 11L,
  control = control_bag(),
  cost = NULL,
  ...
)

# S3 method for class 'formula'
bagger(
  formula,
  data,
  weights = NULL,
  base_model = "CART",
  times = 11L,
  control = control_bag(),
  cost = NULL,
  ...
)

# S3 method for class 'recipe'
bagger(
  x,
  data,
  base_model = "CART",
  times = 11L,
  control = control_bag(),
  cost = NULL,
  ...
)

Arguments

x

A data frame, matrix, or recipe (depending on the method being used).

...

Optional arguments to pass to the base model function.

y

A numeric or factor vector of outcomes. Categorical outcomes (i.e classes) should be represented as factors, not integers.

weights

A numeric vector of non-negative case weights. These values are not used during bootstrap resampling.

base_model

A single character value for the model being bagged. Possible values are "CART", "MARS", "nnet", and "C5.0" (classification only).

times

A single integer greater than 1 for the maximum number of bootstrap samples/ensemble members (some model fits might fail).

control

A list of options generated by control_bag().

cost

A non-negative scale (for two class problems) or a square cost matrix.

formula

An object of class "formula" (or one that can be coerced to that class): a symbolic description of the model to be fitted. Note that this package does not support multivariate outcomes and that, if some predictors are factors, dummy variables will not be created unless by the underlying model function.

data

A data frame containing the variables used in the formula or recipe.

Details

bagger() fits separate models to bootstrap samples. The prediction function for each model object is encoded in an R expression and the original model object is discarded. When making predictions, each prediction formula is evaluated on the new data and aggregated using the mean.

Variable importance scores are calculated using implementations in each package. When requested, the results are in a tibble with column names term (the predictor), value (the importance score), and used (the percentage of times that the variable was in the prediction equation).

The models can be fit in parallel using the future package. The enable parallelism, use the future::plan() function to declare how the computations should be distributed. Note that this will almost certainly multiply the memory requirements required to fit the models.

For neural networks, variable importance is calculated using the method of Garson described in Gevrey et al (2003)

References

Gevrey, M., Dimopoulos, I., and Lek, S. (2003). Review and comparison of methods to study the contribution of variables in artificial neural network models. Ecological Modelling, 160(3), 249-264.

Examples

if (rlang::is_installed(c("recipes", "modeldata"))) {
  library(recipes)
  library(dplyr)

  data(biomass, package = "modeldata")

  biomass_tr <-
    biomass %>%
    dplyr::filter(dataset == "Training") %>%
    dplyr::select(-dataset, -sample)

  biomass_te <-
    biomass %>%
    dplyr::filter(dataset == "Testing") %>%
    dplyr::select(-dataset, -sample)

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

  ctrl <- control_bag(var_imp = TRUE)

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

  # `times` is low to make the examples run faster


  set.seed(7687)
  cart_bag <- bagger(x = biomass_tr[, -6], y = biomass_tr$HHV,
                     base_model = "CART", times = 5, control = ctrl)
  cart_bag

  # ------------------------------------------------------------------------------
  # Other interfaces

  # Recipes can be used
  biomass_rec <-
    recipe(HHV ~ ., data = biomass_tr) %>%
    step_pca(all_predictors())

  set.seed(7687)
  cart_pca_bag <- bagger(biomass_rec, data = biomass_tr, base_model = "CART",
                         times = 5, control = ctrl)

  cart_pca_bag
}
#> Loading required package: dplyr
#> 
#> Attaching package: ‘dplyr’
#> The following objects are masked from ‘package:stats’:
#> 
#>     filter, lag
#> The following objects are masked from ‘package:base’:
#> 
#>     intersect, setdiff, setequal, union
#> 
#> Attaching package: ‘recipes’
#> The following object is masked from ‘package:stats’:
#> 
#>     step
#> Bagged CART (regression with 5 members)
#> 
#> Variable importance scores include:
#> 
#> # A tibble: 5 × 4
#>   term  value std.error  used
#>   <chr> <dbl>     <dbl> <int>
#> 1 PC2   4500.     245.      5
#> 2 PC1   3559.     156.      5
#> 3 PC3   1107.     210.      5
#> 4 PC5    648.     137.      5
#> 5 PC4    468.      71.6     5
#>