Skip to content
Quantoros

XGBoost has 128 features and all three colsample ratios are 0.5. Roughly how many candidates remain per split?

Classic MLLevel: IntermediateXGBoost · Parameters

XGBoost has 128 features and all three colsample ratios are 0.5. Roughly how many candidates remain per split?

Hints
  • Follow the candidate subset through three stages.
  • Multiply the retained fractions.

Answer

About 128 × 0.5³ = 16. Tree, level and node sampling operate cumulatively, subject to rounding and constraints; the three options are not interchangeable aliases.

Source and licence: XGBoost · Parameters · Apache-2.0 · answer: Quantoros editorial

Related questions