Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Anchors
A local explanation method that provides simple and sufficient decision rules (anchors) that probabilistically guarantee the same prediction for a neighborhood of the observation, offering more stable interpretation than LIME.
Shapley Values
The fundamental theoretical concept of SHAP, representing the average marginal contribution of a feature across all possible feature coalitions in a model, ensuring a fair distribution of importance.
Local Surrogate Explanation
An approach that trains a simple and interpretable model (such as a decision tree or linear regression) to approximate the behavior of a complex model only in the restricted neighborhood of a specific prediction.
Observation Neighborhood
The data space defined around a specific observation, used by local interpretation methods to generate variations and approximate the model's behavior in this restricted region.
Local Fidelity
A metric evaluating the accuracy with which a local explanation (such as a surrogate model) reproduces the original model's predictions in the neighborhood of the explained observation.
TreeSHAP
A variant of the SHAP algorithm optimized for decision tree-based models, capable of calculating exact Shapley values much faster by leveraging the intrinsic structure of these models.
KernelSHAP
A SHAP implementation that uses a weighting function (kernel) to estimate Shapley values approximately, making it applicable to any model in an agnostic manner but with higher computational cost.
DeepSHAP
An adaptation of SHAP specifically designed for deep learning models, which combines Shapley values with backpropagation techniques to efficiently compute feature attributions.
Ad-Hoc Explanation
A locally generated explanation specifically for a single instance, without claiming generalization, unlike global explanations that seek to describe the model's overall behavior.
Local Feature Influence
The measure of the impact of a specific feature on the prediction of a single observation, quantifying how varying this feature would change the model's outcome for this specific case.
Individual Prediction Diagnosis
The complete process of analyzing a single prediction using various local methods (LIME, SHAP, counterfactuals) to understand the underlying mechanisms, validate the decision, and identify potential biases.
Local Explanation Stability
The property of a local interpretation method to produce consistent explanations for very similar observations, a critical issue for the trust and reliability of individual diagnostics.
Integrated Gradients
A local attribution method for differentiable models that calculates the importance of a feature by integrating the gradient of the output with respect to that feature along a path from a baseline to the input.
Baseline
A reference point (often a zero vector or average instance) used in attribution methods like integrated gradients to measure the contribution of a feature relative to a neutral or expected state.