Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Anchors
A method that identifies simple and sufficient decision rules (anchors) that explain a model's prediction for a given instance with high fidelity.
Shapley Value
A concept from game theory that measures the average marginal contribution of a player (feature) across all possible coalitions, serving as the foundation for SHAP.
Input Perturbation
The process of creating slight variations in the input data to observe the effect on the model's prediction, used by methods like LIME to build a local neighborhood.
Fidelity
A metric evaluating how faithfully a local explanation (like LIME's simple model) mimics the behavior of the complex model in its neighborhood.
Model-Agnostic Explanation
An interpretability approach that treats the predictive model as a black box, interacting only with its inputs and outputs to generate explanations.
Saliency Map
A visualization that highlights the pixels or features of an input that most influenced a model's prediction, often obtained by computing the gradient.
Kernel Neighborhood
In LIME, a function that defines the proximity between the original instance and the perturbed instances, weighting their influence in the local explanation model.
Explanation Rule
A simple logical condition (e.g., IF feature_A > X AND feature_B < Y) that captures the primary reason for a specific prediction, typical of methods like Anchors.
Post-hoc Interpretability
The analysis of a model after its training to understand its decisions, as opposed to intrinsically interpretable models.
SHAP Kernel Explainer
A SHAP implementation using kernel weighting to estimate Shapley values, making it model-agnostic but potentially slower.
SHAP Tree Explainer
An optimized SHAP algorithm that calculates exact Shapley values for tree-based models (like XGBoost, LightGBM) very efficiently.
Local Surrogate Explanation
The fundamental principle of LIME, consisting of training a simple and interpretable model (surrogate) to approximate the behavior of the complex model locally.