What trade-off does subsample below one introduce in classic gradient boosting?
Classic MLLevel: IntermediateRow subsampling in boosting
What trade-off does subsample below one introduce in classic gradient boosting?
Hints
- Randomize the stage's fitting sample.
- Noise can regularize but also weaken updates.
Answer
Each stage uses a random subset of rows, introducing stochasticity that can reduce variance and computation but increase bias. Tune it jointly with learning rate and stage count.
Source and licence: scikit-learn 1.7.2 — ensemble · BSD-3-Clause · answer: Quantoros editorial