The scikit-learn estimator suite

One fitted RBIG flow wears four identities — it is simultaneously a density model, a generator, an information estimator, and a feature transformer. Every estimator in this suite is a thin, check_estimator- compliant wrapper that exposes one of those identities through the scikit-learn API you already know.

The four identities of a fitted RBIG

Identity	Core method	What it gives you
Density	score_samples(X)	exact log p(x) via change-of-variables
Generator	sample(n)	draws z ~ N(0, I) and inverts the flow
Information	estimate_mi, estimate_tc, …	entropy / MI / TC through total-correlation reduction
Transformer	transform / inverse_transform	an invertible map to a Gaussianized latent space

Because the same fitted object serves all four, the estimators below compose these identities rather than reimplementing anything.

The suite

Estimator	sklearn role	Built on	Example
RBIGOutlierDetector	OutlierMixin	density	19
RBIGReducer	TransformerMixin	information (negentropy)	20
RBIGMISelector	SelectorMixin	information (MI)	21
RBIGKMeans	ClusterMixin	transformer + inverse	22
RBIGBayesClassifier	ClassifierMixin	density (one per class)	23
RBIGFairTransformer	TransformerMixin	transformer + inverse (transport)	24
ResidualDiagnostics	meta RegressorMixin	information (residual MI)	25

All estimators pass scikit-learn's parametrize_with_checks suite and work inside Pipeline and GridSearchCV.

Where they win — and where they don't

The benchmark report compares every estimator against its standard baseline on the same task, including the rows where the baseline wins. In brief:

• MI — for genuinely joint, multivariate MI there is no sklearn baseline; for scalar–scalar screening, KSG is competitive and often more accurate.
• Outliers — competitive-to-best on low-dimensional curved shapes; IsolationForest is the more robust default as dimensionality grows.
• Clustering — a large win on elongated clusters; raw k-means wins on well-separated axis-aligned modes (which is why n_layers_rbig defaults small).
• Classification — nonlinear boundaries (rings, bananas) where LDA is at chance; on genuinely Gaussian classes it matches LDA, no better.
• Preprocessing — pays off most for a linear downstream model on curved data; for an RBF-SVC the choice barely matters.
• Fairness — transport/conditional remove second-order leakage that a linear projection cannot.

See the individual notebooks for the mechanism behind each result.