Using Stratified Sampling to Improve LIME Image Explanations
Published in AAAI Conference on Artificial Intelligence, 2024
We investigate the use of a stratified sampling approach for LIME Image, a popular model-agnostic explainable AI method for computer vision tasks, in order to reduce the artifacts generated by typical Monte Carlo sampling. Such artifacts are due to the undersampling of the dependent variable in the synthetic neighborhood around the image being explained, which may result in inadequate explanations due to the impossibility of fitting a linear regressor on the sampled data. We then highlight a connection with the Shapley theory, where similar arguments about undersampling and sample relevance were suggested in the past. We derive all the formulas and adjustment factors required for an unbiased stratified sampling estimator. Experiments show the efficacy of the proposed approach..
Paper contribution
In this paper we:
- investigate the distribution of the dependent variable in the sampled synthetic neighborhood of LIME Image, identifying in the undersampling a cause that results in inadequate explanations;
- delve into the causes of the synthetic neighborhood inadequacy, recognizing a link with the Shapley theory;
- reformulate the synthetic neighborhood generation using an unbiased stratified sampling strategy;
- provide empirical proofs of the advantage of using stratified sampling for LIME Image on a popular dataset.
Method Availability
- The method is available under: https://github.com/rashidrao-pk/lime_stratified
- Examples and full results given at Lime_stratified_examples
- Links: GitHub, Paper PDF on Arxiv Proceedings of AAAI-24
Datasets and Models
- Dataset: ImageNet
- Model: Resnet-50
How LIME_Image Works
Keywords
XAI · LIME · Stratified Sampling . ML: Transparent, Interpretable, Explainable ML, RU: Stochastic Optimization, SO: Sampling/Simulation-based Search
Authors
| Sr. No. | Author Name | Affiliation | Google Scholar |
|---|---|---|---|
| 1. | Muhammad Rashid | University of Torino, Computer Science Department, C.so Svizzera 185, 10149 Torino, Italy | Muhammad Rashid |
| 2. | Elvio G. Amparore | University of Torino, Computer Science Department, C.so Svizzera 185, 10149 Torino, Italy | Elvio G. Amparore |
| 3. | Enrico Ferrari | Rulex Innovation Labs, Rulex Inc., Via Felice Romani 9, 16122 Genova, Italy | Enrico Ferrari |
| 4. | Damiano Verda | Rulex Innovation Labs, Rulex Inc., Via Felice Romani 9, 16122 Genova, Italy | Damiano Verda |
Citation
If you use our proposed strategy, please cite us:
Rashid, M., Amparore, E. G., Ferrari, E., & Verda, D. (2024). Using Stratified Sampling to Improve LIME Image Explanations. Proceedings of the AAAI Conference on Artificial Intelligence, 38(13), 14785-14792. https://doi.org/10.1609/aaai.v38i13.29397
Recommended citation: Rashid,Muhammad et al. (2024). "." Proceedings of the AAAI Conference on Artificial Intelligence. 38(13).
Download Paper | Download Slides