Local Interpretable Model-agnostic Explanations

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

Definition

• Proximity weight is how close a ouput datum is to the point of interest

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

LIME Usage

• We have an initial model that classifies blue from grey
• Scatter a bunch of points
• Weight each point differently depending on how close to point of interest
• Train a new model for that point - this one looks like a grid to determine the blues and greys

Problems

• XAI Insincerity
• XAI Underlying Mechanism Unexplainability