Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Local Explainability
Level of interpretation aiming to explain a model's prediction for a specific observation, isolating the impact of each feature for that particular case.
Global Explainability
Level of interpretation that aggregates local explanations over a dataset to understand the general behavior of the model and the average importance of features.
Base Value
In SHAP, the average prediction of the model on the training set, serving as the starting point to which feature contributions are added to obtain the final prediction.
Additivity
Fundamental property of SHAP guaranteeing that the sum of SHAP values of all features, added to the base value, equals exactly the model's prediction for an observation.
Efficiency
Axiom of Shapley value stating that the contributions of all features must add up to equal the difference between the model's prediction and the base value.
Symmetry
Axiom of Shapley value ensuring that two features contributing identically to all possible coalitions receive the same importance value.
Dummy
Axiom of Shapley value guaranteeing that a feature providing no marginal contribution to any coalition will be assigned a zero importance value.
Interaction Value
Extension of SHAP values that quantifies the interaction effect between a pair of features, measuring how their combined impact differs from the sum of their individual impacts.