Local Interpretable Model-agnostic Explanations

A form of Model Agnostic Interpretability Finding the local behavior around a prediction of a model.

Process

1. Data perturbation - lime generates dataset of perturbed samples
2. Prediction collection - blackbox model makes predictions on perturbed samples
3. Weight assignment - sample closer to original instance receive higher weights
4. Local model training - interpretable model (like Linear Regression) is trained on weighted dataset
5. Explanation generation - coefficients of local model serve as explanations for original prediction