LIME Stratified Sampling
Published:
This project improves the stability of LIME image explanations by introducing a stratified sampling strategy for perturbation generation.
Overview
LIME is widely used for model-agnostic explanations, but its explanations can vary because of random perturbation sampling. This project addresses that issue by using a more structured sampling strategy to reduce explanation variance.
Main Contributions
- Improved perturbation sampling for LIME image explanations.
- Reduced randomness in local explanation generation.
- More stable and reliable visual explanations.
- Reproducible implementation and example experiments.
Impact
The method helps make LIME-based explanations more consistent, which is important when explanations are used for model debugging, decision support, or trust assessment.
Technologies
Python · LIME · Scikit-learn · Computer Vision · Explainable AI