LIME

LIME (Local Interpretable Model-agnostic Explanations) explains an individual prediction of any machine learning model by fitting a simple, interpretable surrogate model to perturbed samples around that prediction.

LIME(국소적 해석가능 모형불가지 설명)은 특정 예측 주변에서 교란한 표본에 단순하고 해석 가능한 대리모형을 적합하여 어떤 머신러닝 모형의 개별 예측이든 설명하는 방법이다.

How do I explain why a machine learning model made a specific prediction?

머신러닝 모형이 특정 예측을 왜 내렸는지 어떻게 설명하나요?

LIME perturbs the input around one case, records how the black-box model's predictions change, and fits a weighted linear model whose coefficients show which features pushed that prediction up or down. It fits case-level explanations for any classifier or regressor, including tabular, text, and image models.

When to use LIME

• Explaining a single prediction of a black-box classifier or regressor
• Communicating the drivers of an individual decision, such as a credit or risk score
• Checking whether a model relies on sensible features for specific cases

When to use something else

• Consistent, additive attributions with theoretical guarantees are needed: use SHAP values
• A global summary of feature influence across all cases is needed: use permutation importance or partial dependence plots

Assumptions and requirements

• A trained predictive model that returns predictions or probabilities
• A sensible perturbation scheme and kernel width for the data type
• Explanations are local and can be unstable across runs, so a random seed and repeated runs help

What the analysis reports

• Feature weights for each explained prediction
• Bar chart of the top features pushing the prediction up or down
• Local surrogate fit (R²) for each explanation
• Predicted value or class probability being explained

How to report it

LIME explanations indicated that [feature] contributed most to the predicted [class] for [case] (weight = [value]), with a local surrogate R² of [value].

Related methods

• SHAP Values
• Permutation Feature Importance
• Partial Dependence Plot

Run LIME on Quorant