KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Data Perturbation
Process of generating synthetic samples by slightly modifying the characteristics of an original instance to build the local neighborhood required for explanation.
Instance Neighborhood
Set of data points generated by perturbation around a specific observation, used to learn the local behavior of the complex model.
Kernel LIME
LIME variant using a kernel function to define similarity between instances, allowing handling of different data types and representation spaces.
Textual Explanation
Specific application of LIME for natural language processing models, identifying influential words or word groups in text classification.
Image Explanation
Use of LIME for computer vision, where superpixels (image segments) are perturbed to determine key visual regions in a prediction.
Local Decision Boundary
Boundary approximated by the surrogate model in the neighborhood of an instance, representing how the complex model separates classes locally.
Surrogate Model Complexity
Parameter controlling the sophistication of the simple model used in LIME, balancing interpretability and fidelity to local approximation.
Approximation Bias
Potential error introduced by the fact that the surrogate model does not perfectly capture the behavior of the original model except in a limited area.
Explanation Function
Mathematical function that, for a given instance, returns an interpretable explanation by minimizing the fidelity loss between the original model and the surrogate.