chimeraboost
pronounced kai-MEER-uh-boost
Lightning-fast Gradient Boosting, near-Catboost quality, all in Python
📖 Documentation: bbstats.github.io/chimeraboost
#2 On TabArena GBDT Elo (Defaults)
Installation
pip install chimeraboost
Quickstart
from chimeraboost import ChimeraBoostClassifier, ChimeraBoostRegressor
# quality picks the speed/accuracy trade-off: 1 is fastest - 5 is strongest,
# default is 3.
clf = ChimeraBoostClassifier(quality=4)
clf.fit(X, y, cat_features=[0, 1], sample_weight=w)
proba = clf.predict_proba(X_test)
reg = ChimeraBoostRegressor()
reg.fit(X, y)
What?
- Regression, quantile regression, binary and multiclass classification
- Fast training and inference, all in python
- Extremely high quality predictions
How?
- Bagging as a first-class feature (using
n_ensembles) - Automatic early stopping
- Automatic linear-leaf auditioning
- Replay mechanism for faster refitting after early stopping
- Gradient-matrix free multi-quantile split search
- numba is very fast
Why?
- I want to be able to modify my GBDT library at will
- I know Python and I don't know C
Documentation
Benchmarking
Want to help? Run either one, then open an issue or PR with the JSON.
pip install -e ".[bench,competitors]"
python benchmarks/run_benchmarks.py --synth --seeds 3 --save # quick, synthetic
python benchmarks/run_benchmarks.py --decide --seeds 3 --save # slower, 103 real datasets
Each writes benchmarks/results/<timestamp>.json:
{
"provenance": {"chimeraboost": "0.30.0", "platform": "Linux-6.1", "cpu_count": 12,
"libraries": {"catboost": "1.2.10", "lightgbm": "4.6.0"}},
"records": [
{"dataset": "diabetes", "model": "ChimeraBoost", "seed": 0,
"metrics": {"primary": -59.82, "rmse": 59.82}, "fit_time": 0.23}
]
}
Inspirations / Citations
Most ideas in ChimeraBoost were someone else's.
- CatBoost, Prokhorenkova et al., NeurIPS 2018. Oblivious trees (the tree type itself from Kohavi & Li, IJCAI 1995), ordered target statistics, feature combinations, ordered boosting, and the size-dependent automatic learning rate (
adaptive_learning_rate). - XGBoost, Chen & Guestrin, KDD 2016. Regularized objective, second-order split gain, Newton leaf estimation,
min_child_weight. - LightGBM, Ke et al., NeurIPS 2017. Histogram-based split finding, which that paper itself treats as prior art (McRank, Li et al. 2007; pGBRT, Tyree et al. 2011).
- Minimum Variance Sampling, Ibragimov & Gusev, NeurIPS 2019. Gradient-weighted row sampling (
subsample). - Quantized GBDT training, Shi, Ke et al., NeurIPS 2022. Integer gradient histograms (
quantize_gradients). - SketchBoost, Iosipoi & Vakhrushev, NeurIPS 2022, and GBDT-MO, Zhang & Jung. Vector leaves and the projected gradient used for multiclass and multi-quantile splits.
- Linear-leaf trees, Shi, Li & Li, IJCAI 2019 (arXiv:1802.05640). Piece-wise-linear regression trees (
linear_leaves); model trees back to Quinlan's M5, 1992. - TreeSHAP, Lundberg et al., Nature Machine Intelligence 2020 (orig. SHAP, NeurIPS 2017). Exact additive feature attributions (
shap_values). - OpenFE, Zhang et al., ICML 2023 (arXiv:2211.12507). Automated pairwise feature generation (
cross_features).
1 maintainer
Metadata
1 maintainer
Release files for chimeraboost 0.33.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| chimeraboost-0.33.0.tar.gz | 332.5 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| chimeraboost-0.33.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 513.6 kB
Release files / chimeraboost-0.33.0.tar.gz
| Download URL | chimeraboost-0.33.0.tar.gz |
|---|---|
| Size | 332.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e980585c38733474ce044102388ee502e40e4771d576c1e8a5f9675a3451725f
|
|
BLAKE2b-256 checksum How to use checksums |
f259a6055406a5af7c34d96f03378ccf31cda73e748e2c2dfbea0774f8eac09b
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.13.13
|
Release files / chimeraboost-0.33.0-py3-none-any.whl
| Download URL | chimeraboost-0.33.0-py3-none-any.whl |
|---|---|
| Size | 181.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
9537b80e63b2de9a3311b03821a9bec755e304873ef8b35e50379297e6b8d779
|
|
BLAKE2b-256 checksum How to use checksums |
3c6482f01fae0858bfd34df9e06eee3b3b7c382d3c087b0e556460a5ac1e6697
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.13.13
|
