Why can reusing a constructed LightGBM Dataset distort a search over mindatainleaf?
Classic MLLevel: IntermediateLightGBM · Parameters
Why can reusing a constructed LightGBM Dataset distort a search over min_data_in_leaf?
Hints
- Some decisions happen before boosting starts.
- A removed feature cannot be reconsidered later.
Answer
Feature prefiltering can remove unsplittable columns when the Dataset is constructed under an earlier threshold. Rebuild the Dataset for the candidate settings or disable prefiltering deliberately; the latter may increase training cost.
Source and licence: LightGBM · Parameters · MIT · answer: Quantoros editorial