uxplain

uxplain explains which features make a model more or less uncertain. It wraps scikit-learn compatible models with conformal prediction and uses SHAP, PDP/ICE or LIME to explain summaries of their prediction intervals or sets.

The package supports regression, through crepes and conformalized quantile regression (CQR), and classification, through crepes. A single pipeline handles model fitting, calibration and explanation.

Installation

Requires Python 3.10 or later.

pip install uxplain

Regression

Explain prediction interval width with SHAP. fit() automatically reserves a separate calibration sample.

from sklearn.datasets import make_regression
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split
from uxplain import UncertaintyExplanationPipeline

X, y = make_regression(n_samples=300, n_features=4, noise=15, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

pipe = UncertaintyExplanationPipeline(
    model=RandomForestRegressor(n_estimators=50, random_state=42),
    confidence=0.9,
    xai_method="shap",
    uncertainty_metric="width",
    random_state=42,
)
pipe.fit(X_train, y_train)
result = pipe.explain(X_test[:5], show_plots=False)

print(result.interval_width)
print(result.explanation_values)

Classification

Use a classifier and set_size to explain how many labels enter the prediction set.

from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier

X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

pipe = UncertaintyExplanationPipeline(
    model=RandomForestClassifier(n_estimators=50, random_state=42),
    uncertainty_metric="set_size",
    random_state=42,
)
pipe.fit(X_train, y_train)
result = pipe.explain(X_test[:5], show_plots=False)

print(result.prediction_set)
print(result.explanation_values)

Learn more

The getting-started guide covers calibration, other uncertainty summaries and explainer options. See the notebooks for more examples and the release guide for development and publishing.

Coverage guarantees depend on the conformal method's statistical assumptions; they do not extend to the explanations themselves.

Authors

• Tomas Rodriguez Taborda — Universidad Nacional de Colombia, Medellín
• Veronica Seguro Varela — Universidad Nacional de Colombia, Medellín
• Rafael Izbicki — Federal University of São Carlos
• Johnatan Cardona Jimenez — Universidad Nacional de Colombia, Medellín

License

MIT.