v0.3.0

Added

• Survival analysis: NestedCVSurvival with Cox PH models, Harrell's and Uno's concordance indices, integrated Brier score and coefficient stability. Install with pip install nestkit[survival].
• make_survival_target to build a survival target from event and duration arrays.
• Missing values are accepted in X: an imputer placed in a Pipeline is fitted inside every fold.
• explainability extra for SHAP importances.
• Tutorial notebook on survival analysis.

Changed

• matplotlib and seaborn are now core dependencies, and matplotlib 3.10 or newer is required.

Fixed

• Binary targets whose labels are not 0 and 1 are now supported: predictions, conformal sets and results_ come back in the original labels, and, as in scikit-learn, classes_[1] is the positive class. This also repairs isotonic calibration, threshold selection with Youden's J and the expected calibration error, which were all wrong on such targets.
• Mondrian prediction intervals fell short of their target coverage on small bins. Bins too small for the requested alpha are now merged, and a bound the sample cannot determine is returned as infinite.
• SHAP importances for Pipeline estimators were computed on the raw features instead of the preprocessed ones.
• plot_comparison raised TypeError on matplotlib 3.9 and newer.
• A fold whose roc_auc could not be computed no longer turns the whole roc_auc summary row into NaN.
• Comparing models from a single fold now reports that a variance needs at least two folds, instead of returning NaN with numpy warnings.
• The feature-selection stability index reported a perfect 1.0 when a fold's importances were all equal; it now returns NaN with a warning.
• Corrected the documented example for holm_bonferroni_correction.