Research map / Probabilistic and scientific modeling / Molecules and proteins
Selective 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 selective explanations that detect low-quality amortized Shapley explanations via learned/deep uncertainty and refine the flagged ones with an optimally weighted combination of amortized and Monte Carlo explanations.
Problem
Amortized feature attribution explainers are cheap (one inference) but can produce inaccurate explanations for individual inputs, while Monte Carlo explainers are accurate but costly per input.
Approach
Selectively route inputs: uncertainty metrics (deep ensemble variance and a learned MLP uncertainty predictor) flag points where the amortized explainer is likely poor; for those, an "explanation with initial guess" linearly combines the amortized and Monte Carlo explanations with a bin-wise closed-form weight calibrated on held-out data.
Key results
- Learned uncertainty selects low-quality amortized explanations almost as accurately as the oracle on the large models (Yelp, Toxigen); e.g., for 50% coverage average MSE drops from 0.11 to 0.07 on Yelp and from 0.032 to 0.007 on Toxigen with theoretical guarantees.
- Selective explanations with only 20% of points receiving initial-guess refinement increase Spearman's correlation on the worst-performing points, e.g., on Yelp the worst 4% beat the amortized explainer's worst 10%.
- On Toxigen, selective explanations with about 1000 inferences outperform SVS-50 using about 6000 inferences.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select the lowest-uncertainty half of explanations for amortized output | Amortized explainer · UCI-Adult | MSE: 0.014 -> 0.003 | Helped |
| Select the lowest-uncertainty half of explanations for amortized output | Amortized explainer · UCI-News | MSE: 0.032 -> 0.025 | Helped |
| Select the lowest-uncertainty half of explanations for amortized output | Amortized explainer · Yelp Review | MSE: 0.11 -> 0.07 | Helped |
| Select the lowest-uncertainty half of explanations for amortized output | Amortized explainer · Toxigen | MSE: 0.032 -> 0.007 | Helped |
| Combine the amortized explanation with a Monte Carlo explanation using the optimized initial-guess weight instead of using Monte Carlo alone (ablation) | Selective explanations · UCI-Adult, UCI-News, Yelp Review, and Toxigen | MSE | Helped |
| Combine the amortized explanation with a Monte Carlo explanation using the optimized initial-guess weight instead of using Monte Carlo alone (ablation) | Selective explanations · UCI-News and Toxigen | MSE | Helped |
| Provide explanations with initial guess for 20% of inputs | Selective explanations · UCI-Adult, UCI-News, Yelp Review, and Toxigen | MSE | Helped |
| Provide explanations with initial guess for 20% of the population using SVS-12 | Selective explanations · Yelp Review | Spearman's correlation: from 0.2 to more than 0.3 | Helped |
| Provide explanations with initial guess for 20% of inputs | Selective explanations · Yelp Review | Spearman's correlation | Helped |
| Provide explanations with initial guess for 60% of points and amortized explanations for the remainder | Selective explanations · Yelp Review | local fidelity | Helped |
| Increase the quality of the Monte Carlo explainer used for recourse | Selective explanations · UCI-Adult, UCI-News, Yelp Review, and Toxigen | MSE | Helped |
| Use selective explanations rather than vanilla Monte Carlo time sharing | Monte Carlo time sharing · Toxigen | MSE | Helped |
Surprising findings
- Naive use of a Monte Carlo explanation for low-quality points can worsen MSE when the Monte Carlo explanation is noisier than the amortized one, while the initial-guess combination still reduces MSE.
Limitations
- Empirical results focus only on Shapley values; applicability to other attribution methods is left for future work.
- Evaluation limited to text classification with LLMs and tabular classifiers; image classifiers not explored.
- High-quality reference explanations require very many inferences, so evaluation uses only 4000 samples per dataset.
Experimental setup
Datasets: UCI-Adult UCI-News Yelp Review Toxigen
Benchmarks: UCI-Adult UCI-News Yelp Review Toxigen
Metrics: MSE to high-quality Shapley values Spearman's correlation model accuracy under token removal (perturbation curve)
Baselines: Amortized explainer (stochastic amortization) Amortized SHAP for LLMs Kernel SHAP Shapley Value Sampling Naive Monte Carlo recourse (lambda=0) Oracle uncertainty selector
Builds on: Shapley Value Sampling Kernel SHAP Stochastic amortization Amortized SHAP for LLMs Deep ensembles
Closest papers by meaning
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.89)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.88)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.86)
- GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation (ICML 2025 · similarity 0.86)
- Stochastic Amortization: A Unified Approach to Accelerate Feature and Data Attribution (NeurIPS 2024 · similarity 0.86)
- Saliency strikes back: How filtering out high frequencies improves white-box explanations (ICML 2024 · 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.