Research map / Learning theory and algorithms / Prediction and decision losses
Probabilistic Stability Guarantees for Feature Attributions
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.
Introduces soft stability (stability rate under random additive feature perturbations) and a model-agnostic sampling certifier (SCA) that certifies hard and soft stability, plus a Boolean-analysis account of why mild MuS random-masking smoothing can help.
Problem
Certifying the robustness of feature attributions to adding features is hard: existing hard-stability certificates (MuS) rely on heavily smoothed classifiers and give conservative, quickly vacuous guarantees.
Approach
Relax hard stability to a probabilistic stability rate estimated by uniform sampling of additive perturbations (SCA), with Hoeffding-based guarantees and no smoothing required; analyze smoothing via the monotone (unanimity) Boolean basis, where random masking contracts each coefficient by lambda^|T|.
Key results
- SCA gives non-trivial soft stability guarantees at larger radii, whereas MuS-based hard-stability certificates become vacuous beyond radius 1/(2*lambda).
- LIME and SHAP yield higher stability rates than gradient-based methods on vision models; all methods beat the random baseline.
- Mild smoothing (lambda=0.5) generally improves stability while preserving accuracy, with transformers (ViT, RoBERTa) showing a more gradual accuracy decline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCA instead of MuS for stability certification | Feature attribution stability certification · Vision and language tasks | stability guarantee strength | Helped |
| use LIME or SHAP instead of gradient-based attribution methods | Vision models · ImageNet subset | stability rate | Helped |
| use feature attribution methods instead of random attribution | Vision models · ImageNet subset | stability rate | Helped |
| compare explanation methods on RoBERTa | RoBERTa · TweetEval subsets | stability rate | No clear effect |
| apply mild random-mask smoothing (λ ≥ 0.5) | Vision and language classifiers · ImageNet subset and TweetEval | stability rate | Helped |
| apply mild random-mask smoothing (λ ≥ 0.5) | Vision and language classifiers · ImageNet subset and TweetEval | accuracy | Hurt |
| decrease the MuS smoothing parameter λ | MuS-smoothed classifiers · Vision and language tasks | MuS stability guarantee strength | Helped |
| decrease the MuS smoothing parameter λ | MuS-smoothed classifiers · Vision and language tasks | accuracy | Hurt |
| use mild smoothing rather than heavy smoothing | ViT and RoBERTa · ImageNet subset and TweetEval | accuracy | Helped |
Surprising findings
- Mild smoothing can improve stability rates even though SCA does not require smoothing.
- Explanation methods are less distinguishable on RoBERTa than on vision models.
Limitations
- Smoothing-based guarantees (MuS) only hold for radii up to 1/(2 lambda) and degrade accuracy as lambda shrinks.
- Soft stability is a one-sided, additive-perturbation notion and does not cover removal of features.
Experimental setup
Datasets: ImageNet (2000-image subset) TweetEval (emoji, emotion, hate, irony, offensive, sentiment)
Benchmarks: ImageNet subset TweetEval
Metrics: stability rate accuracy certified radius
Baselines: MuS LIME SHAP Integrated Gradients MFABA random attribution
Builds on: MuS random masking smoothing Hoeffding inequality LIME SHAP
Closest papers by meaning
- Stability Guarantees for Feature Attributions with Multiplicative Smoothing (NeurIPS 2023 · similarity 0.92)
- Feature Bagging Provides Stability (ICML 2026 · similarity 0.88)
- Pixel-level Certified Explanations via Randomized Smoothing (ICML 2025 · similarity 0.87)
- Certified Robust Interpretability via Concept-Space Stability under Interventional Proxies (NeurIPS 2026 · similarity 0.85)
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.85)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.85)
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.