Why does minchildweight=20 not require twenty observations in every logistic XGBoost leaf?
Classic MLLevel: IntermediateXGBoost · Parameters
Why does min_child_weight=20 not require twenty observations in every logistic XGBoost leaf?
Hints
- What quantity is accumulated?
- For logistic loss, curvature depends on p.
Answer
The threshold is the sum of Hessians, not the row count. Logistic curvature depends on current probabilities and sample weights; inspect effective curvature before interpreting the parameter as a sample minimum.
Source and licence: XGBoost · Parameters · Apache-2.0 · answer: Quantoros editorial