FaithfulnessAuditor<T>

Models & Types · AiDotNet.Interpretability

Audits how faithful a per-feature attribution vector is to a model, using perturbation curves. This is the check mainstream explainability libraries omit: they produce attributions but do not measure whether removing the “important” features actually changes the model’s output.

Constructors

Constructor Summary
FaithfulnessAuditor(int maxSteps, double topFraction) Creates a faithfulness auditor.

Methods

Method Summary
FaithfulnessReport<T> Audit(Func<Vector<T>, T> score, Matrix<T> data, Vector<T> attributions) Audits attributions against score (a scalar model output for an input) over the rows of data.