Why can feature scaling matter in LightGBM lineartree even when ordinary constant-leaf trees are mostly insensitive to…
Classic MLLevel: IntermediateLightGBM · Parameters
Why can feature scaling matter in LightGBM linear_tree even when ordinary constant-leaf trees are mostly insensitive to units?
Hints
- The leaf is no longer a single constant.
- Linear fitting has numerical conditioning concerns.
Answer
Linear leaves fit numerical models within leaves, so conditioning and scale affect that fit and its penalties. Scale appropriately and validate the linear-tree assumptions; the usual threshold-tree intuition is incomplete here.
Source and licence: LightGBM · Parameters · MIT · answer: Quantoros editorial