LIME — Local Interpretability

The Key Difference: Local vs Global

SHAP: Global Credit Assignment

• Shapley values apply game theory
• Guarantees mathematical correctness
• Slower computationally
• Best for: Understanding why a model made a specific decision

LIME: Local Approximation

• Approximates complex model with simple linear model locally
• Model-agnostic (works with ANY model)
• Fast and computationally cheap
• Best for: Quick explanations for non-technical users

How LIME Works

1. Perturb: Add noise to the input (slightly change features)
2. Predict: Run all perturbed inputs through the black-box model
3. Weight: Weight perturbed samples by distance from original
4. Fit: Fit a simple linear model (Logistic Regression) on perturbed data
5. Explain: The linear model coefficients are LIME's explanation!

Intuition:

"I can't understand why this complex neural network said 'cat', so I'll fit a simple line around that prediction and see which features mattered."

LIME Force Plot Example

Original prediction: Classifier says "DOG" (probability 0.95)

LIME approximation using a simple rule:

• Feature "has_fur" = strong positive contributor
• Feature "has_tail" = moderate positive contributor
• Feature "meows" = strong negative contributor