External Libraries Based on scikit-learn

yobx.sklearn.to_onnx() converts fitted estimators to ONNX from xgboost, lightgbm, category_encoders, and imbalanced-learn using the same registry-based architecture as the other yobx.sklearn converters.

Package	Module
category_encoders	yobx.sklearn.category_encoders
xgboost	yobx.sklearn.xgboost
lightgbm	yobx.sklearn.lightgbm
imbalanced-learn	yobx.sklearn.imblearn

Comparison: XGBoost vs LightGBM

Both XGBoost and LightGBM are gradient-boosted tree libraries — their fitted models consist of an ensemble of binary decision trees. The converters map these trees to ONNX TreeEnsemble* operators and share the same high-level structure, differing only in how split conditions are expressed and how the raw margin is post-processed. The table below summarises the key differences between the converters for xgboost and lightgbm.

Property	XGBoost	LightGBM
Tree dump method	get_dump(dump_format='json')	booster_.dump_model()
Split direction (numerical)	x < threshold (BRANCH_LT)	x ≤ threshold (BRANCH_LEQ)
Categorical splits	Not supported by converter	Expanded to BRANCH_EQ chains
Base-score bias	Added from base_score cfg	None (baked into leaf values)
Regression transform	Identity / Sigmoid / Exp	Identity / Exp
Multi-class targets per round	n_classes trees	n_classes trees
Binary classifier raw outputs	1 tree per round (sigmoid)	1 tree per round (sigmoid)

imbalanced-learn

imbalanced-learn extends scikit-learn with resampling techniques for imbalanced datasets. Resampling only happens at training time, so ONNX inference pipelines skip any step that exposes fit_resample and only convert the remaining transformers and the final estimator.

See yobx.sklearn.imblearn for the full API reference.