AI-woordenlijst
Het complete woordenboek van kunstmatige intelligentie
Counterfactual
Hypothetical instance that minimally modifies input features to change an AI model's prediction, providing an intuitive explanation of the model's behavior.
Counterfactual space
Mathematical domain containing all possible modifications of input features that can reverse the model's decision, often explored by optimization algorithms.
Counterfactual distance
Metric quantifying the gap between the original instance and its counterfactual version, essential for ensuring plausible and interpretable explanations for users.
Counterfactual validity
Fundamental criterion verifying that the generated counterfactual scenario actually produces the desired model prediction, ensuring the reliability of the explanation.
Counterfactual proximity
Principle stating that a counterfactual should be as close as possible to the original instance to be considered a relevant and understandable explanation.
Counterfactual plausibility
Evaluation of the realism of a counterfactual scenario in the real world, crucial for explanations to be acceptable and usable by decision-makers.
Counterfactual actionability
Measure of the extent to which the modifications suggested by a counterfactual are feasible and controllable by the user, determining its practical value as a decision-making tool.
Minimal counterfactual
Counterfactual explanation that modifies the smallest possible number of features while changing the model's prediction, optimizing simplicity and interpretability.
Counterfactual Generation
Algorithmic process of creating hypothetical scenarios that reverse the model's decision, often using constrained optimization techniques.
Counterfactual Optimization
Mathematical approach aimed at finding the optimal counterfactual by minimizing a trade-off between distance, validity, and plausibility according to predefined objectives.
Counterfactual Robustness
Ability of a counterfactual explanation to remain valid in the face of slight variations in the model or data, ensuring the stability of generated recommendations.
Multiple Counterfactuals
Set of several counterfactual scenarios offering different paths to modify the model's prediction, allowing users to choose among various action options.
Counterfactual Criticality
Analysis of the relative importance of modified features in a counterfactual, identifying the most influential factors for changing the model's decision.
Minimal Perturbation
Fundamental principle in counterfactual generation aiming to modify input data as little as possible while achieving the desired prediction change.
Realism Constraint
Set of rules imposed during counterfactual generation to ensure that produced scenarios respect the physical, logical, or domain-specific laws of the problem.
Feature Space
Multidimensional domain in which input data evolves, serving as a framework for exploring and generating valid and relevant counterfactuals.
Post-hoc explanation
Interpretation method applied after model training, to which counterfactual generation belongs, to explain decisions without modifying the algorithm.
Similarity metric
Mathematical function used to quantify the resemblance between the original instance and its counterfactual, essential for evaluating the relevance of generated explanations.