Research map / Probabilistic and scientific modeling / Variational and generative sampling
Using Stratified Sampling to Improve LIME Image Explanations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes stratified sampling with adjustment factors for LIME Image's synthetic neighborhood, motivated by undersampling of the dependent variable and links to Shapley sampling theory.
Problem
LIME Image's Monte Carlo (Bernoulli 0.5) sampling of superpixel masks undersamples masks with few or many preserved superpixels, so the dependent variable distribution is flattened and the linear surrogate's coefficients become near-uniform and uninformative for images with many superpixels.
Approach
Replace the fixed Bernoulli(0.5) mask sampling with stratified sampling where each mask's Bernoulli parameter q is drawn uniformly, so all k+1 mask-count strata are equally represented, and correct the resulting bias with a per-stratum adjustment factor (k+1)·C(k,i)/2^k multiplied into LIME's kernel weights.
Key results
- Across 150 ImageNet images with k=50 to 200 superpixels, stratified sampling keeps the dependent-variable range well covered, whereas Monte Carlo produces low-CV explanations for some images, especially at k=200.
- Stratified sampling yields better R^2 than Monte Carlo on average, per the comparison plots (no exact figures given in the text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use stratified sampling with adjustment factors instead of Monte Carlo sampling | LIME Image · ImageNet Object Localization dataset, 150 images | RC | Helped |
| Use stratified sampling with adjustment factors instead of Monte Carlo sampling | LIME Image · ImageNet Object Localization dataset, 150 images | CV of feature attributions | Helped |
| Use stratified sampling with adjustment factors instead of Monte Carlo sampling | LIME Image · ImageNet Object Localization dataset, 150 images | R² | Helped |
| Use stratified sampling with adjustment factors instead of Monte Carlo sampling | LIME Image · Images with sufficient sampled distribution | explanation consistency | No clear effect |
Experimental setup
Datasets: ImageNet Object Localization dataset
Benchmarks: ImageNet Object Localization dataset
Metrics: coefficient of variation of beta (CV) range coverage of Y (RC) R^2
Baselines: LIME Image with Monte Carlo sampling
Builds on: LIME Shapley value quick shift
Closest papers by meaning
- SLICE: Stabilized LIME for Consistent Explanations for Image Classification (CVPR 2024 · similarity 0.90)
- REPEAT: Improving Uncertainty Estimation in Representation Learning Explainability (AAAI 2025 · similarity 0.84)
- I-CEE: Tailoring Explanations of Image Classification Models to User Expertise (AAAI 2024 · similarity 0.83)
- Explainable Saliency: Articulating Reasoning with Contextual Prioritization (CVPR 2025 · similarity 0.83)
- GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation (ICML 2025 · similarity 0.82)
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.