Research map / Learning theory and algorithms / Prediction and decision losses
CatBoost: unbiased boosting with categorical features
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CatBoost Ordered rather than LightGBM | gradient-boosted trees with ordered target statistics · Adult, Amazon, Click, Epsilon, Appetency, Churn, Internet, Upselling, Kick | logloss: +2.4%; +17%; +1.2%; +1.5%; +0.4%; +0.1%; +6.8%; +0.3%; +3.5% | Helped |
| Use CatBoost Ordered rather than XGBoost | gradient-boosted trees with ordered target statistics · Adult, Amazon, Click, Epsilon, Appetency, Churn, Internet, Upselling, Kick | logloss: +2.2%; +17%; +1.2%; +11%; +0.4%; +0.5%; +7.9%; +0.04%; +3.2% | Helped |
| Replace ordered boosting with Plain boosting (ablation) | CatBoost · Adult and Internet | logloss: +1.1%; +3.9% | Hurt |
| Replace ordered boosting with Plain boosting (ablation) | CatBoost · Amazon, Click, Churn, Kick | logloss: -0.6%; -0.05%; -0.06%; -0.2% | No clear effect |
| Replace ordered boosting with Plain boosting (ablation) | CatBoost · Epsilon, Appetency, Upselling | logloss: +0.6%; +0.5%; +0.1% | Hurt |
| Replace ordered target statistics with greedy target statistics (ablation) | CatBoost Ordered · Amazon, Click, Appetency, Churn, Internet, Upselling, Kick | logloss: +40%; +13%; +24%; +12%; +33%; +57%; +22% | Hurt |
| Replace ordered target statistics with holdout target statistics (ablation) | CatBoost Ordered · Adult, Amazon, Click, Appetency, Churn, Internet, Upselling, Kick | logloss: +2.1%; +8.3%; +1.5%; +1.6%; +0.9%; +2.6%; +1.6%; +1.3% | Hurt |
| Replace ordered target statistics with leave-one-out target statistics (ablation) | CatBoost Ordered · Adult, Amazon, Click, Appetency, Churn, Internet, Upselling, Kick | logloss: +5.5%; +4.5%; +2.7%; +8.5%; +1.6%; +27%; +3.9%; +3.7% | Hurt |
| Allow categorical feature combinations up to 2 rather than 1 (ablation) | CatBoost · average across evaluated datasets | logloss: 1.86% | Helped |
| Increase the number of permutations from 1 to 3 or 9 (ablation) | CatBoost · average across evaluated datasets | logloss: 0.19%; 0.38% | Helped |
Closest papers by meaning
- NGBoost: Natural Gradient Boosting for Probabilistic Prediction (ICML 2020 · similarity 0.87)
- The Many Faces of Optimal Weak-to-Strong Learning (NeurIPS 2024 · similarity 0.85)
- A Boosting-Type Convergence Result for AdaBoost.MH with Factorized Multi-Class Classifiers (NeurIPS 2024 · similarity 0.84)
- Wasserstein Gradient Boosting: A Framework for Distribution-Valued Supervised Learning (NeurIPS 2024 · similarity 0.84)
- Robust Minimax Boosting with Performance Guarantees (NeurIPS 2025 · similarity 0.83)
- Revisiting Agnostic Boosting (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.