FaithfulnessReport<T>
Models & Types · AiDotNet.Interpretability
The faithfulness of a feature-attribution explanation to a model: do the features it highlights actually drive the model’s output? Reported as deletion/insertion AUC and ERASER-style comprehensiveness/sufficiency, plus a single normalized score.
Constructors
| Constructor | Summary |
|---|---|
FaithfulnessReport() |
Properties
| Property | Summary |
|---|---|
T Comprehensiveness { get; set; } |
Comprehensiveness (ERASER): output drop when the top-attributed features are removed. |
T DeletionAuc { get; set; } |
Deletion AUC (Most-Relevant-First): mean model output as top-attributed features are progressively removed. |
T FaithfulnessScore { get; set; } |
A single normalized faithfulness score in roughly [0, 1] combining the above (insertion minus deletion, plus comprehensiveness minus sufficiency). |
T InsertionAuc { get; set; } |
Insertion AUC: mean model output as top-attributed features are progressively inserted into a baseline. |
T Sufficiency { get; set; } |
Sufficiency (ERASER): output drop when ONLY the top-attributed features are kept. |
int SampleCount { get; set; } |
Number of rows the audit was averaged over. |