Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Selectivity
Measures the ability of an explanation to focus on the most relevant features for a specific prediction, ignoring noise or non-informative variables. It quantifies the degree of sparsity of the explanation.
R-squared Fidelity Score
Coefficient of determination used to evaluate the fidelity of a linear explanation model by measuring the proportion of the variance in the black-box model's predictions that is explained by the explanation model.
Area Under the Precision-Recall Curve (AUPRC) for Interpretability
Adaptation of the AUPRC metric to evaluate the quality of feature-based explanations, treating the selection of the most important features as a binary classification problem.
Explanation Computational Cost
Quantifies the time and hardware resources required to generate an explanation. This is an essential practical criterion for deploying interpretability methods at scale.
Discontinuity Rate
Metric evaluating the sudden, non-linear variation of explanations for very similar data instances. A high rate indicates an unstable and potentially non-intuitive explanation.
Sensitivity Score
Quantifies how explanations change in response to modifications of the underlying model (e.g., after retraining). Low sensitivity is desirable for consistent explanations.
Coherence Metric
Evaluates whether explanations for different predictions are logically consistent with each other, for example, if features identified as important for one class are oppositely important for an antagonistic class.
Fidelity Loss
Objective function used when training interpretable models that penalizes the discrepancy between the predictions of the interpretable model and those of the black-box model it seeks to explain.
Sparsity Index
Measures the ratio of non-zero features or coefficients in an explanation relative to the total number of possible features. A high index indicates a simpler and potentially more easily interpretable explanation.