After adding monotonic constraints, hist trees become unexpectedly shallow. Why might larger maxbin help?
Classic MLLevel: IntermediateXGBoost · Monotonic constraints
After adding monotonic constraints, hist trees become unexpectedly shallow. Why might larger max_bin help?
Hints
- Two restrictions are acting on candidate splits.
- Increase resolution, not just maximum depth.
Answer
Histogram binning limits candidate thresholds, and monotonicity can reject all remaining candidates. More bins may expose feasible splits, at a memory/time cost; it does not guarantee better validation quality.
Source and licence: XGBoost · Monotonic constraints · Apache-2.0 · answer: Quantoros editorial