API reference¶

The top-level API measures instability for arbitrary fitted procedures and includes one representative-model selector plus fixed-design linear calibration tools. Inclusion means the implementation and workflow are supported; it does not imply a universal stabilization guarantee.

Bootstrap audits¶

`bootstrap_predictions`	Refit a model on bootstrap resamples and return every prediction it made.
`bootstrap_instability`	Measure how much a model's predictions move when the training data is perturbed.

bootstrap_predictions returns every refitted prediction and the per-case statistics. bootstrap_instability returns aggregate summaries and Monte Carlo standard errors.

Model-selection frontier¶

`stability_frontier`	Sweep a parameter grid and return the validation-score/instability tradeoff.
`pareto_front`	Keep the configurations no other configuration beats on both axes.

The score used to construct a frontier is a validation score. It is not a final test-set performance estimate.

Tree-structure audits¶

`split_features`	Return the multiset of features tested down to a given depth.
`split_feature_paths`	Return the feature sequence each row is tested against on its way to a leaf.
`explanation_instability`	How much the structure changes across independently fitted trees.
`root_agreement`	Fraction of fits that test the most common root feature.
`path_agreement`	Fraction of rows routed through the same feature sequence by most fits.

Read structural instability beside prediction instability. A consistently shallow or inaccurate tree can have perfectly stable structure.

Plots¶

Install plotting dependencies with pip install "stable-cart[plots]".

`plot_prediction_instability`	Draw the instability plot: original prediction against resampled predictions.
`plot_mape_by_prediction`	Show instability as a function of predicted value: who the model is unsure about.
`plot_stability_frontier`	Plot one or more model families on the validation-score/stability plane.

Supported estimator and analytic tools¶

RepresentativeEstimator supports multiclass classification and selects a single fitted candidate by validation-set prediction centrality. Its observed stability benefit is task dependent. The linear functions are exact or calibrated calculations under their documented fixed-design assumptions; they are not general estimators.

`RepresentativeEstimator`	Select one fitted estimator closest to a candidate-pool centroid.
`linear_instability`	Analytic prediction instability of least squares, conditional on the design.
`linear_frontier`	Trace an oracle or plug-in frontier for fixed-design linear prediction.
`shrinkage_coefficients`	Coefficients of the estimator that achieves the frontier point at `mu`.