Main API¶

`kfc_procedure.KFCProcedure` ¶

Bases: BaseEstimator

Full KFC pipeline estimator (sklearn-compatible).

This estimator implements the complete KFCProcedure framework:

K-step → F-step → C-step

It performs divergence-aware clustering, cluster-wise local model fitting, and final ensemble aggregation.

Parameters:

Name	Type	Description	Default
`divergences`	`list`	List of Bregman divergences used in K-step clustering. Can include strings or BaseBregmanDivergence instances.	required
`local_model`	`str or BaseLocalModel`	Base learner used in F-step for cluster-wise modeling.	required
`combiner`	`str or BaseCombiner`	Aggregation strategy used in C-step.	required
`divergences_params`	`dict`	Optional parameters for each divergence.	`None`
`local_model_params`	`dict`	Parameters for local models in F-step.	`None`
`combiner_params`	`dict`	Parameters for aggregation strategy in C-step.	`None`
`task`	`('regression', 'classification')`	Learning task type.	`"regression"`
`n_clusters`	`int`	Number of clusters per divergence.	`3`
`max_iter`	`int`	Maximum iterations for KMeans optimization.	`300`
`tol`	`float`	Convergence tolerance for clustering.	`1e-4`
`verbose`	`int`	Enable debugging output. Levels: 0 - silent 1 - basic info 2 - detailed debug 3 - trace (per-iteration / per-cluster)	`0`
`random_state`	`int or None`	Random seed for reproducibility.	`None`

Attributes:

Name	Type	Description
`kstep_`	`KStep`	Fitted divergence-aware clustering stage.
`fstep_`	`FStep`	Fitted local model learning stage.
`cstep_`	`CStep`	Fitted aggregation stage.

Methods:

Name	Description
`fit`	Fit full KFC pipeline.
`predict`	Predict using full pipeline.
`predict_proba`	Predict probabilities (classification only).

Notes

The training procedure uses a train/test split internally:

- K-step + F-step trained on training split
- C-step trained on held-out predictions

This reduces overfitting in the aggregation stage and improves generalization of the ensemble combiner.