chimeraboost 0.14.1

CatBoost-inspired gradient boosting in pure Python with a numba backend

chimeraboost

What if CatBoost was slightly worse, 12× faster, and all in Python?

📖 Documentation: bbstats.github.io/chimeraboost

• Installation

pip install chimeraboost

• Cold start

  The first fit() in a fresh environment JIT-compiles the numba kernels — a few seconds, machine-dependent. The compiled kernels are cached on disk (cache=True), so every later session in the same environment skips this and starts fast.

• Sample code:

from chimeraboost import ChimeraBoostClassifier, ChimeraBoostRegressor

# classification
clf = ChimeraBoostClassifier(early_stopping=True, n_ensembles=2)
clf.fit(X, y, cat_features=[0, 1], sample_weight=w)
proba = clf.predict_proba(X_test)

# regression (RMSE, MAE, or Quantile)
reg = ChimeraBoostRegressor(loss="Quantile", alpha=0.9, early_stopping=True, n_ensembles=10)
reg.fit(X, y)

• What?

  • Exceedingly opinionated GBDT library that only depends on common Python libraries
    • Categorical features (catboost-like processing) and sample weights
    • Bagging as a first-class feature (n_ensembles)
    • Automatic early stopping, with optional grouped splitting for the validation set
  • Supports regression, quantile regression, binary and multiclass classification
  • Exact SHAP explanations (model.shap_values(X)) — interventional TreeSHAP computed exactly (not sampled) thanks to the oblivious tree structure, with the linear-leaf slopes included

• Tuning tips

  • Interaction-heavy regression: raise depth to 8–10 (default 6 is conservative to protect small data).

• Inspirations / Citations

  • XGBoost — Chen & Guestrin, KDD 2016 — regularized objective, Newton leaf estimation, column subsampling
  • LightGBM — Ke et al., NeurIPS 2017 — histogram-based split finding
  • CatBoost — Prokhorenkova et al., NeurIPS 2018 — ordered boosting, ordered target statistics, oblivious trees
  • Linear-leaf trees — Shi et al., IJCAI 2019 (arXiv:1802.05640) — piece-wise-linear regression trees (linear_leaves)
  • TreeSHAP — Lundberg et al., Nature Machine Intelligence 2020 (orig. SHAP, NeurIPS 2017) — exact additive feature attributions (shap_values)
  • Hierarchical shrinkage — Agarwal et al., ICML 2022 (arXiv:2202.00858) — leaf regularization (hs_lambda)
  • TabArena — Erickson et al., NeurIPS 2025 (arXiv:2506.16791) — tabular benchmark used for evaluation

• Why?

  • I want to be able to modify my GBDT library at will
  • I know Python and I don't know C