Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Explanation Fidelity
Metric quantifying the correspondence between the black box model's predictions and those of the interpretable model used to generate explanations. High fidelity indicates that the explanation faithfully represents the original model's behavior in the considered local region.
Explanation Stability
Indicator assessing the consistency of explanations generated for similar instances or for the same instance with slight variations. Stability ensures that explanations do not vary erratically in response to minor changes in input data.
Explanatory Completeness
Metric measuring an explanation's ability to capture all relevant factors influencing the model's decision. A complete explanation should integrate all significant features without omitting crucial elements.
Semantic Relevance
Evaluation of the consistency between the generated explanation and domain knowledge or expected human reasoning. This metric quantifies whether the produced explanations align with domain experts' logic and intuition.
Explanation Compactness
Measure of an explanation's conciseness, assessing the ratio between the amount of information provided and its structural complexity. A compact explanation prioritizes the most relevant elements while minimizing informational redundancy.
Explanatory Robustness
Ability of an explanation to maintain its validity in the face of perturbations or adversarial attacks on input data. This metric evaluates the resistance of explanations to malicious manipulations aimed at misleading.
Granularity Level
Level of detail at which an explanation operates, ranging from global explanations (entire model) to local explanations (specific instance). Granularity determines the precision and specificity of the provided interpretation.
Inter-explanation Consistency
Metric evaluating the logical consistency between different explanations generated for varied but semantically similar instances. This metric ensures that explanations follow reasonable and non-contradictory patterns.
Perceived Usability
Qualitative indicator measuring the ease with which users can understand, interpret, and apply the generated explanations. Perceived usability assesses the adequacy between the technical complexity of the explanation and the user's cognitive abilities.
Explanatory Verifiability
Ability to independently confirm or refute the validity of the explanations provided by the model. Verifiability allows users to validate the consistency of explanations against external knowledge or empirical tests.
Explanation Gap
Quantitative difference between the inherent complexity of the model and the simplicity of its explanation. A high gap may indicate significant information loss during the explanatory simplification process.
Causal Specificity
Measure assessing whether an explanation correctly identifies cause-effect relationships rather than mere correlations. Causal specificity distinguishes factors that actually influence the decision from those that are merely co-occurring.
Explanatory Generalization
Ability of a local explanation to apply consistently to other similar instances in the dataset. This metric evaluates whether the identified explanatory patterns can be extrapolated beyond the specific case studied.
Explanatory Confidence
Quantified level of certainty associated with an explanation, indicating the probability that the explanation is correct. Explanatory confidence allows users to assess the reliability of the interpretations provided by the system.
Explanation Fairness
Metric evaluating whether the generated explanations treat different demographic groups or subpopulations fairly. Explanation fairness ensures the absence of discriminatory bias in how decisions are justified.
Explanatory Coverage
Proportion of the feature space or instances for which the model can generate valid explanations. High coverage ensures that the explanation system can operate on the majority of cases encountered in practice.
Explanatory Latency
Computational time required to generate an explanation after the model has produced its prediction. This metric is crucial for real-time applications where explanations must be provided quickly.
Counterfactual Fidelity
Specific measure evaluating the quality of counterfactual explanations in terms of the minimality of required changes and the plausibility of generated scenarios. This metric ensures that the proposed counterfactuals are realistic and actionable.