Tree-based algorithms do not require feature scaling before fitting, and I am working on gradient boosted tree models (and random forest) without scaling features.
I'm curious if feature scaling affects the feature importance in a meaningful way?
To be specific,
- When both X (explanatory) and Y are NOT scaled
- When X (explanatory) is scaled and Y are not scaled
- When both X and Y are scaled
If it does affect, what is the reason for that?