Skip to content
Quantoros

Why can scaleposweight improve rare-class retrieval yet make raw probabilities unsuitable as population probabilities?

Classic MLLevel: IntermediateXGBoost · Parameters

Why can scale_pos_weight improve rare-class retrieval yet make raw probabilities unsuitable as population probabilities?

Hints
  • Training weights change the loss being optimized.
  • Ranking and calibration measure different properties.

Answer

The weighted objective changes the relative cost of positive errors and generally the fitted probability target. Check calibration on the intended population and choose a decision threshold separately; improved ranking does not prove calibrated probabilities.

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

Related questions