Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Reliable Post hoc Explanations: Modeling Uncertainty in Explainability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add posterior credible intervals to LIME-style feature importance estimates | BayesLIME · COMPAS, German Credit, ImageNet, and MNIST | 95% credible interval coverage: 95.5; 96.9; 91.4–95.8; 95.2–97.2 | Helped |
| Add posterior credible intervals to KernelSHAP-style feature importance estimates | BayesSHAP · COMPAS, German Credit, ImageNet, and MNIST | 95% credible interval coverage: 87.9–94.6; 89.2 to 98.4% | Mixed |
| Use focused sampling instead of random perturbation sampling | BayesLIME · ImageNet French Bulldog | model queries to stabilize error uncertainty: within a couple hundred vs over 1,000 | Helped |
| Use focused sampling instead of random sampling to approach the reference explanation | BayesLIME · ImageNet | queries to reach L1 distance 0.1: 300 queries vs upwards of 450 queries; 30-40% | Helped |
| Use Bayesian variants rather than LIME/SHAP | BayesLIME and BayesSHAP · COMPAS, German Credit, MNIST Digit 4, and ImageNet French Bulldog | stability: 53% | Mixed |
| Select the explanation using P(ε=0) rather than fidelity | BayesLIME · MNIST user study | user error rate: 25.7% -> 30.7% | Hurt |
| Use closed-form Bayesian inference rather than Bayesian mixture regression with MCMC | BayesLIME and BayesSHAP · ImageNet, one instance | explanation time: 139.2 seconds -> 20.3 seconds; 139.2 seconds -> 21.1 seconds | Helped |
| Set strong uncertainty priors instead of uninformative priors (ablation) | BayesLIME · MNIST Digit 4 | credible interval calibration | Hurt |
| Use focused sampling rather than random sampling (ablation) | BayesLIME · ImageNet | L1 distance to reference explanation: 0.1 at 300 queries vs upwards of 450 queries; around 0.06 at ~1000 | Helped |
Closest papers by meaning
- Improving Perturbation-based Explanations by Understanding the Role of Uncertainty Calibration (NeurIPS 2025 · similarity 0.90)
- Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations (NeurIPS 2022 · similarity 0.90)
- Trust Regions for Explanations via Black-Box Probabilistic Certification (ICML 2024 · similarity 0.88)
- Dissenting Explanations: Leveraging Disagreement to Reduce Model Overreliance (AAAI 2024 · similarity 0.87)
- Evaluate with the Inverse: Efficient Approximation of Latent Explanation Quality Distribution (AAAI 2025 · similarity 0.86)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.86)
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.