Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Counterfactual
Minimal and modified data instance compared to an original case, which changes the model's prediction to a desired output, serving to explain the model's borderline decision.
Counterfactual Explanation
Interpretability method that explains a prediction by presenting a hypothetical scenario (counterfactual) where the model's decision would have been different, thus clarifying the decision criteria.
Counterfactual Validity
Criterion ensuring that a generated counterfactual indeed produces the expected alternative prediction from the model, guaranteeing the reliability and relevance of the provided explanation.
Counterfactual Proximity
Measure of the distance between the original instance and the counterfactual, often quantified by a norm (e.g., L1, L2), aiming to ensure that the explanation is plausible and easily interpretable.
Counterfactual Sparsity
Principle stating that a counterfactual should modify the smallest possible number of features of the original instance to maximize the clarity and actionability of the explanation.
Counterfactual Plausibility
Evaluation of the credibility of a counterfactual in the real world, ensuring that the suggested modifications are feasible and do not correspond to an aberrant or impossible instance.
Counterfactual Diversity
Objective to generate a set of counterfactuals that are not redundant, offering several distinct alternative paths to achieve a different prediction and thus enriching the understanding of the model.
Counterfactual Cost
Function weighting the modifications of features in a counterfactual, reflecting the difficulty or cost (monetary, temporal, etc.) of implementing these changes in reality.
Causal Counterfactuals
Advanced approach where counterexamples are generated while respecting causal relationships between variables, ensuring that proposed scenarios do not violate real-world constraints.
Counterfactual Robustness
Ability of a counterexample to maintain its alternative prediction in the face of slight variations or noise, indicating the stability of the model's decision boundary in that region.
Adversarial Counterfactual Generation
Use of adversarial learning techniques to create counterexamples, often for security or auditing purposes, to test model vulnerabilities and weaknesses.
Counterfactual Latent Space
Method that searches for counterexamples in a lower-dimensional representation space (latent space) to improve computational efficiency and consistency of generated instances.
Counterfactual Optimization Methods
Set of algorithms (e.g., constraint programming, gradient descent) used to solve the problem of finding the optimal counterexample by minimizing a loss function combining proximity and validity.
Multi-class Counterfactual Explanations
Extension of counterexamples to classification problems with more than two classes, where instances are generated to switch to any other target class, not just the opposite class.