Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Fidelity
The extent to which an explanation faithfully reflects the model's internal reasoning, evaluating whether the explanation's predictions match those of the model on perturbed data.
Comprehensibility
A subjective or objective measure of how easily a human can understand an explanation, often related to the complexity of the explanation model (e.g., number of rules, depth of a tree).
Sufficiency
The ability of a subset of features, identified by an explanation, to maintain the model's original prediction, indicating that these features are sufficient to justify the decision.
Necessity
Evaluates whether the absence of a feature (or set of features) identified as important by the explanation significantly changes the model's prediction.
Causal Inference Score (CIS)
A metric quantifying an explanation's ability to identify actual causal relationships rather than mere correlations, by testing the effects of interventions on variables.
Explanation Robustness
Measures the variation in explanations when the model or input data undergo adversarial attacks or noise, assessing the interpretation's resistance to manipulation.
Feature Coherence
Evaluates whether the features deemed important by an explanation are semantically or logically coherent with each other, enhancing the plausibility of the overall explanation.
Selectivity Rate
An indicator measuring the proportion of features or rules used by an explanation relative to the total available, favoring parsimonious explanations.
Relevance Function
Mathematical function that quantifies the contribution of a feature or set of features to the model's final prediction, serving as the basis for many interpretability metrics.
Inter-Annotator Agreement
Statistical measure (e.g., Cohen's Kappa score) assessing the level of consensus among different human experts on the quality or correctness of an explanation, validating its subjectivity.
Confirmation Bias
Metric evaluating whether an explanation only reinforces the user's pre-existing beliefs without challenging the model, measuring the risk of fallacious interpretations.
Discriminative Power
Ability of an explanation to clearly distinguish features that positively influence the prediction from those that negatively influence it, improving interpretation clarity.
Global Fidelity
Evaluates an explanation's ability to faithfully represent the model's overall behavior across the entire data space, often at the expense of local accuracy.
Counterfactual Score
Metric assessing the quality of a counterfactual explanation based on the minimal perturbation required to change the model's prediction and the plausibility of the generated scenario.
Semantic Depth
Measures the level of abstraction of an explanation, quantifying whether it is based on low-level features (pixels) or higher-level concepts (objects, ideas) that are more intelligible.