Research map / Architectures and efficiency / Quantization
Quantized Training of Gradient Boosting Decision Trees
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 3-bit stochastic rounding with leaf-value refitting instead of full-precision LightGBM training | GBDT · Kitsune | AUC: 0.950561 -> 0.951309 | Helped |
| Use 3-bit stochastic rounding with leaf-value refitting instead of full-precision LightGBM training | GBDT · Year | RMSE: 8.956278 -> 8.937374 | Helped |
| Use 3-bit stochastic rounding with leaf-value refitting instead of full-precision LightGBM training | GBDT · Criteo | AUC: 0.803791 -> 0.803768 | No clear effect |
| Use 2-bit quantized training instead of LightGBM+ | GPU LightGBM+ · Epsilon | GPU total time (seconds): 87.12 -> 39.04 | Helped |
| Replace stochastic rounding with round-to-nearest (ablation) | 2-bit quantized GBDT with refitting · Higgs | AUC: 0.845587 -> 0.795991 | Hurt |
| Remove leaf-value refitting (ablation) | 3-bit SR quantized GBDT · Year | RMSE: 8.937374 -> 8.990034 | Hurt |
| Remove leaf-value refitting (ablation) | 3-bit SR quantized GBDT · LETOR | NDCG@10: 0.522220 -> 0.507689 | Hurt |
| Pack gradient and hessian accumulations together (ablation) | CPU quantized GBDT · Criteo | overall training time (seconds): 946 -> 675 | Helped |
| Pack gradient and hessian accumulations together (ablation) | CPU quantized GBDT · Higgs | overall training time (seconds): 156 -> 122 | Helped |
| Increase gradient precision from 2-bit to 5-bit with SR refit (ablation) | Quantized GBDT · Epsilon | AUC: 0.949472 -> 0.950298 | Helped |
| Increase gradient precision from 2-bit to 5-bit with SR refit (ablation) | Quantized GBDT · Kitsune | AUC: 0.952703 -> 0.949229 | Mixed |
Closest papers by meaning
- LightGBM: A Highly Efficient Gradient Boosting Decision Tree (NeurIPS 2017 · similarity 0.91)
- SketchBoost: Fast Gradient Boosted Decision Tree for Multioutput Problems (NeurIPS 2022 · similarity 0.90)
- Gradient Boosted Decision Trees for High Dimensional Sparse Output (ICML 2017 · similarity 0.89)
- Gradient Regularized Newton Boosting Trees with Global Convergence (NeurIPS 2026 · similarity 0.89)
- NRGBoost: Energy-Based Generative Boosted Trees (ICLR 2025 · similarity 0.86)
- GradTree: Learning Axis-Aligned Decision Trees with Gradient Descent (AAAI 2024 · similarity 0.85)
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.