User Overview

This section explains how to use kfc-procedure without reading the source code.

Goal

KFCProcedure helps train models on heterogeneous data. It assumes that one global model may not be ideal, so it creates a clusterwise learning pipeline:

Step	Name	Meaning
K-step	Clustering	Split observations into similar groups using Bregman divergences
F-step	Fitting	Train local supervised models in each cluster
C-step	Combining	Combine divergence-level predictions into one final prediction

When to use it

Use KFCProcedure when your dataset may contain hidden subgroups, one model underfits different regions of the data, or you want a local-model ensemble with a scikit-learn-style interface.

Small datasets

Avoid large n_clusters when the dataset is small. Some clusters may have too few samples for stable local training.

Safe first choices

Task	Estimator	Divergence	Local model	Combiner
Regression	KFCRegressor	euclidean	linear_regression	mean
Classification	KFCClassifier	euclidean	decision_tree_classifier	majority_vote

Main registered components

Component type	Available values
Divergences	euclidean, gkl, is, logistic
Regression combiners	mean, weighted_mean, stacking_regressor, gradientcobra, mixcobra
Classification combiners	majority_vote, stacking_classifier, combined_classifier
Local models	many sklearn estimators auto-registered in snake_case