LIME Paper Flashcards
What is LIME (Local Interpretable Model-agnostic Explanations)?
LIME is a model-agnostic technique that explains individual predictions by approximating the complex model locally with an interpretable linear model.
How does LIME explain a prediction?
LIME perturbs the input data to generate similar instances and learns a linear model that approximates the black-box model in the vicinity of the instance being explained.
What is local fidelity in LIME?
Local fidelity ensures that the explanation provided by LIME accurately represents how the model behaves around the specific instance being explained.
How does LIME select interpretable data representations?
LIME creates simplified representations of the data, such as binary vectors indicating the presence or absence of words in text classification or superpixels in image classification.
What are the main steps of the LIME algorithm?
- Perturb the input data. 2. Learn an interpretable model for perturbed data. 3. Use the interpretable model to provide a local explanation.
What is SP-LIME (Submodular Pick LIME)?
SP-LIME selects a representative set of explanations for multiple predictions by using submodular optimization to avoid redundancy and maximize the coverage of important features.
What are the desired characteristics of explanation methods according to LIME?
Explanations should be interpretable, locally faithful, and model-agnostic, and they should provide both local explanations for individual predictions and global insights into the model.
How does LIME handle trust issues in models?
LIME helps users identify untrustworthy features and improves trust by explaining predictions and showing how models behave on representative instances.
What are some applications of LIME in model selection?
LIME can be used to compare models with similar accuracy by explaining their predictions and helping practitioners choose models that avoid exploiting untrustworthy features.
