Validate explanations

Validation helps you check whether an attribution vector behaves as expected for a fitted classifier. FPDE provides deletion and insertion perturbation curves for this purpose.

Compute perturbation curves

from fpde import perturbation_curves

attributions, details = engine.explain_one(X_test[0], lambda_hyb=0.5)

curves = perturbation_curves(
    model,
    X_test[0],
    attributions,
    details["target_label"],
    engine.baseline,
    fractions=(0.0, 0.1, 0.3, 0.5, 1.0),
)

print(curves["deletion_prob"])
print(curves["insertion_prob"])

Features are ranked by signed positive attribution in descending order. Deletion removes top-ranked evidence for the target class. Insertion restores top-ranked evidence for the target class.

Run practical checks

Confirm predict_proba returns at least two class columns.
Confirm model.classes_ contains the labels used to fit FPDE prototypes.
Confirm every explained row has the same number of features as X_train.
Check that attribution values are finite.
Check that details["target_label"] and details["rival_label"] are different.
Inspect details["exactness_residual"] for numerical stability.

Avoid validation leakage

Use separate data splits for selection and final reporting. For example, fit prototypes on training data, choose lambda_hyb on validation data, and report explanations on test data.

Do not select lambda_hyb on the same examples you use for final claims. That makes the reported validation behavior less informative.

Record enough context

Save the validation settings with the output:

Fractions used in deletion and insertion curves
Baseline vector source
lambda_hyb or lambda_hyb_grid
normalize
anchor_strategy
eps
Model and preprocessing artifacts

⌘I

Getting started

Guides

Concepts

Reference

Changelog

Validate explanations

Compute perturbation curves

Run practical checks

Avoid validation leakage

Record enough context

​Compute perturbation curves

​Run practical checks

​Avoid validation leakage

​Record enough context

Compute perturbation curves

Run practical checks

Avoid validation leakage

Record enough context