A CatBoost LossFunctionChange importance is negative. Is that necessarily a serialization bug?
Classic MLLevel: IntermediateCatBoost · Feature importance
A CatBoost LossFunctionChange importance is negative. Is that necessarily a serialization bug?
Hints
- The quantity measures a change in loss, not a nonnegative split count.
- Approximate removal and retraining differ.
Answer
No. Under its sign convention, approximate feature removal can improve the evaluated metric, producing negative importance. Inspect the evaluation data and approximation; it is not equivalent to a fully retrained ablation.
Source and licence: CatBoost · Feature importance · Apache-2.0 · answer: Quantoros editorial