Compute Weighted Feature Importance from Base Models and Meta-Learner for Stacking Models

This function computes the feature importance by weighing the feature importances from multiple base models in a stacking ensemble, combined with the meta-learner's model importance. The final importance score for each feature is calculated by multiplying the base model's feature importance with the meta-learner's weight for each base model.

Usage

calculate_feature_importance_stacking(
  base_importance,
  base_models,
  meta_learner
)

Arguments

base_importance

A list where each element corresponds to a base model and contains a data frame with feature importances. Each data frame should have a column called importance (either for the positive class or overall, depending on the type of model).

base_models

A character vector with the names of the base models whose feature importances are provided in base_importance.

meta_learner

A caret object representing the trained meta-learner model. This model is used to obtain weights for each base model, based on their performance in the ensemble.

Value

A data frame with two columns:

features

The feature names.

final_importance

The final weighted importance score for each feature, calculated by summing the weighted importances across all base models. Features are sorted in descending order of their final importance score.

Details

The function extracts feature importance values from the base models, then computes the weighted importance for each feature based on the meta-learner's performance. The meta-learner's feature importance is normalized to ensure the sum of the importances across all models is 1, and it is used as the weight for each base model. The feature importances from all base models are then aggregated and weighted by their respective meta-learner importance scores.

See also

varImp