I applied the gradient boosted classification trees algorithm via the mlr3 and gbm packages. My task is classification with 8 predictors. I extracted the importance scores from the learner utilizing the command learner_tuned$importance. If I am not mistaken the method used to derive the relative importance is the one Breiman introduced, which is the sum of the improvement i^2 for the non-terminal nodes where each predictor utilized as splitting variable. I have 2 questions:
- In case of classification task what is this measure? The MCE?
- I got the results depicted in the next Figure. In case the improvement refers to the MCE or Entropy reduction, how is it possible to take scores with values greater than 10? Does the gbm package scales up the scores like rpart package does? Thank you in advance.
