yobx.sklearn.imblearn.balanced_bagging#

yobx.sklearn.imblearn.balanced_bagging.sklearn_balanced_bagging_classifier(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: BalancedBaggingClassifier, X: str, name: str = 'balanced_bagging_classifier') → str | Tuple[str, str][source]#

Converts an imblearn.ensemble.BalancedBaggingClassifier into ONNX.

BalancedBaggingClassifier is a bagging ensemble where each sub-estimator is trained on a balanced bootstrap sample (obtained via random under-sampling). At inference time, the resampling step is inactive; each sub-estimator is therefore effectively a plain classifier.

Probabilities from all sub-estimators are averaged (soft aggregation) and the winning class is determined by an argmax over the averaged probability vector. Each sub-estimator is applied to the feature subset recorded in estimators_features_ (all features by default).

The sub-estimators stored in estimators_ are imblearn.pipeline.Pipeline instances wrapping a resampler and a classifier. This converter dispatches to the registered converter for imblearn.pipeline.Pipeline (defined in yobx.sklearn.imblearn.easy_ensemble), which transparently drops the resampler at inference time.

Graph structure (two sub-estimators as an example):

X ──Gather(cols_0)──[sub-pipeline 0]──► (_, proba_0) (N, C)
X ──Gather(cols_1)──[sub-pipeline 1]──► (_, proba_1) (N, C)
        Unsqueeze(axis=0) ──► proba_0 (1, N, C), proba_1 (1, N, C)
            Concat(axis=0) ──► stacked (E, N, C)
                ReduceMean(axis=0) ──► avg_proba (N, C)
                    ArgMax(axis=1) ──Cast──Gather(classes_) ──► label
Parameters:

  • g – the graph builder to add nodes to

  • sts – shapes and types defined by scikit-learn

  • outputs – desired output tensor names; two entries for (label, probabilities), one entry for label only

  • estimator – a fitted BalancedBaggingClassifier

  • X – name of the input tensor

  • name – prefix used for names of nodes added by this converter

Returns:

label tensor name, or tuple (label, probabilities)