#impurity-importance

Question 1: Single choice

hardAI-checkedDecision Trees & Ensembles

A gradient-boosted tree model is trained on a binary feature and on a continuous feature with thousands of distinct values; both are pure noise, unrelated to the target. Which feature will the impurity-based (split-gain) importance typically rank higher?

Neither; impurity importance gives every noise feature exactly zero1
Both equally, because split-gain importance is normalized by each feature's number of distinct values2
The continuous feature, because its many possible split points let it fit noise more often3
The binary feature, because a single clean split is more informative4

#impurity-importance

Timed test