KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Surrogate Model
A simple and interpretable machine learning model trained to approximate the behavior of a complex model, thereby enabling understanding of the original model's predictions.
Global Surrogate Model
An interpretable model that mimics the global behavior of a black-box model across the entire dataset, providing an overview of the complex model's decisions.
Local Surrogate Model
A simple model that approximates the behavior of a complex model only in a specific neighborhood of an individual prediction, thus explaining decisions at the local level.
Model Fidelity
A measure of a surrogate model's ability to faithfully reproduce the predictions of the original black-box model, often evaluated by the coefficient of determination R² or mean squared error.
Surrogate Decision Tree
A simple decision tree used as a surrogate model to approximate the behavior of a complex model, providing an intuitive visual interpretation of decision rules.
Feature Relevance
A quantitative measure of the relative importance of each input feature in a model's predictions, calculated through the coefficients of the surrogate model or other weighting methods.
Local Linear Regression
A linear surrogate model fitted on a weighted subset of data around a specific prediction, allowing local explanation of the relationships between features and prediction.
Counterfactual Explanations
An approach that generates minimally modified examples to change a model's prediction, often implemented via surrogate models to identify critical features.
Model Complexity
Measure of the structural sophistication of a model, where surrogate models favor low complexity (shallow trees, linear models) to ensure interpretability.
Piecewise Approximation
Strategy where the feature space is divided into regions, each with its own simple surrogate model, allowing flexible approximation while maintaining local interpretability.
Kernel Weighting
Technique used in local surrogate models to give more weight to samples close to the point of interest, ensuring better local approximation of the model's behavior.
Perturbed Sampling
Method of generating synthetic data around a prediction by perturbing the original features, used to train local surrogate models on relevant neighborhoods.
Interpretive Validity
Evaluation criterion that measures whether the explanations provided by a surrogate model are consistent with domain knowledge and useful for human decision-making.
Surrogate Function
Simplified mathematical representation that approximates the complex decision function of the original model, essential for making predictions understandable to non-technical users.