What are monotonic constraints in GBDTs? Give an example where they improve trust or generalization, and explain the tradeoff.
formulate your answer, then —
tldr
Monotonic constraints force predictions to move consistently with selected features. They can improve trust, reduce spurious patterns, and encode domain knowledge in GBDTs. The risk is underfitting or wrong behavior if the true relationship is not monotonic.
follow-up
- How would you test whether a monotonic constraint is appropriate?
- What is the difference between monotonicity and feature importance?
- When might a constraint improve production trust even if offline AUC drops slightly?