Skip to contents

Candidate Formulas Weighted by Feature Importance

Source: R/formula_generation.R
sample_weighted_formulas.Rd

Draws random subsets of features, each feature chosen with probability related to its weight (for example the absolute log hazard ratio of a univariate screen), as candidate model formulas. A thin front end to [sample_candidate_formulas()], which has more constraints (required terms, exclusions, hierarchy).

Usage

sample_weighted_formulas(
  feature_weights,
  n_samples = 25,
  min_terms = 2,
  max_terms = length(feature_weights),
  include_full = TRUE,
  response = NULL,
  seed = NULL,
  data = NULL
)

Arguments

feature_weights

Named numeric vector of non-negative weights (names = features). A feature with weight 0 is never drawn, except in the full formula.

n_samples

Number of distinct formulas.

min_terms, max_terms

Smallest and largest number of features per formula.

include_full

Also include the formula with every feature (added if the draw did not produce it, so the result can have `n_samples + 1` rows).

response

Optional left-hand side as text, for example `"Surv(time, status)"`.

seed

Optional integer; the caller's random-number state is restored.

data

Optional data frame (training rows). Features that are zero-variance columns of it are removed before sampling (also from the full formula), with a message.

Value

The tibble of [sample_candidate_formulas()] (`id`, `formula`, `terms`, `n_terms`).

Details

Compute weights from the **training** data of the current fold only; weights from all rows leak held-out information into the candidate set.

See also

[adaptive_cox_formulas()], [expand_feature_formulas()], [calculate_feature_weights()]

Examples

w <- c(age = 0.8, bmi = 0.2, lactate = 1.1, hr = 0.5)
sample_weighted_formulas(w, n_samples = 4, min_terms = 2, max_terms = 3, seed = 1)
#> # A tibble: 5 × 4
#>      id formula   terms     n_terms
#>   <int> <list>    <list>      <int>
#> 1     1 <formula> <chr [2]>       2
#> 2     2 <formula> <chr [2]>       2
#> 3     3 <formula> <chr [2]>       2
#> 4     4 <formula> <chr [3]>       3
#> 5     5 <formula> <chr [4]>       4