Skip to contents

Setup ConditionalSuperLearner Hyperparameters

Source: R/115_MetaLearners.R
setup_ConditionalSuperLearner.Rd

Setup hyperparameters for the Conditional SuperLearner, which selects the best model from a library conditional on the covariates rather than combining them.

Usage

setup_ConditionalSuperLearner(
  base_learners = list(setup_GLM(), setup_GLMNET(), setup_Ranger()),
  meta_learner = setup_Ranger(),
  inner_resampling_config = setup_Resampler(n_resamples = 10L, type = "KFold"),
  n_iterations = 4L,
  loss = NULL,
  init = "full",
  min_region_size = 10L,
  expand_search_spaces = TRUE,
  ifw = FALSE
)

Arguments

base_learners

List of Hyperparameters objects: The experts. Names label the regions and the reported region sizes; unnamed entries are named after their algorithm.

meta_learner

Hyperparameters object: The oracle, a classifier fitted on the extended dataset.

inner_resampling_config

ResamplerConfig object: Cross-validation scheme, which must hold out every case exactly once.

n_iterations

(Tunable) Integer [1, Inf): Alternations between fitting the oracle and refitting the experts.

loss

Optional Character {"squared_error", "log_loss"}: Per-case loss the oracle minimizes.

init

Character {"full", "random"}: Region each expert starts from.

min_region_size

Integer [1, Inf): Fewest cases an expert's region may hold before the expert keeps its previous fit instead of being refitted.

expand_search_spaces

Logical: If TRUE, expand a base learner's search space into one expert per combination.

ifw

(Tunable) Logical: If TRUE, use Inverse Frequency Weighting in classification.

Value

ConditionalSuperLearnerHyperparameters object.

Details

The model is sum_k 1\{o(x) = k\} F_k(x): an "oracle" o assigns each case to one of the K "experts" F_k, and that expert predicts it. Oracle and experts are fitted by alternating, in the manner of k-means: with the experts fixed, the oracle minimizes the total loss, which reduces to a weighted K-class classification over an extended dataset holding each case once per expert; with the oracle fixed, each expert is refitted on the cases assigned to it. The losses are cross-validated throughout, so an expert is never scored on a case it was trained on.

The oracle is a classifier over the original covariates whatever the outcome is, so with more than two experts it must handle more than two classes. Its variable importance – which covariates decide which model applies – is what get_varimp returns for a fitted model, and a simple oracle such as setup_CART makes the partition itself readable.

Supports regression and binary classification.

Reference: Valdes, Interian, Gennatas & van der Laan, "Conditional Super Learner", IEEE Transactions on Pattern Analysis and Machine Intelligence (2022). doi:10.1109/TPAMI.2021.3131976

Author

EDG

Examples

csl_hyperparams <- setup_ConditionalSuperLearner(
  base_learners = list(setup_GLM(), setup_CART())
)
csl_hyperparams
#> <ConditionalSuperLearnerHyperparameters>
#>         hyperparameters: 
#>                                    base_learners: <chr> GLM, CART
#>                                     meta_learner: <chr> Ranger
#>                          inner_resampling_config: <chr> 10 independent folds
#>                             expand_search_spaces: <lgc> TRUE
#>                                              ifw: <lgc> FALSE
#>                                     n_iterations: <int> 4
#>                                             loss: <NUL> NULL
#>                                             init: <chr> full
#>                                  min_region_size: <int> 10
#> tunable_hyperparameters: <chr> ifw, n_iterations
#>   fixed_hyperparameters: <chr> base_learners, meta_learner, inner_resampling_config, expand_search_spaces, loss, init, min_region_size
#>                   tuned: <int> -1
#>               resampled: <int> 0
#>               n_workers: <int> 1