Research map / Label-efficient and robust learning / Recognition and long-tail
Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse
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 responsiveness scores, the probability that an arbitrary actionable intervention on a feature flips a decision to the target prediction, to highlight features that actually support recourse, and shows SHAP/LIME explanations often highlight unresponsive features or explain fixed predictions.
Problem
Feature-highlighting explanations (SHAP, LIME) used for adverse action notices can provide reasons without recourse: they highlight immutable or unresponsive features and explain predictions that cannot be changed.
Approach
Score each feature by the fraction of its reachable points (under actionability constraints and an intervention model with downstream effects) that attain the target prediction, computed model-agnostically via enumeration or sampling of reachable sets using MIP-based routines.
Key results
- On HELOC with logistic regression, 22.2% of denied individuals receive fixed predictions and 92.7% of LIME explanations and 77.3% of SHAP explanations highlight no responsive feature.
- Responsiveness-based explanations contain only responsive features (100% all-features-responsive across datasets and models), versus 0.2% of action-aware SHAP explanations for logistic regression on HELOC.
- On Give Me Some Credit with XGBoost, 94% of LIME/SHAP-based explanations are unlikely to lead to recourse, falling to 72% with direction and degree of change information.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Feature Responsiveness scores instead of LIME to select highlighted features | Logistic regression · FICO | % all features responsive: 0.0% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of SHAP to select highlighted features | Logistic regression · FICO | % all features responsive: 0.0% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of SHAP actionable to select highlighted features | Logistic regression · FICO | % all features responsive: 0.2% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of LIME to select highlighted features | XGBoost · FICO | % all features responsive: 0.0% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of SHAP to select highlighted features | XGBoost · FICO | % all features responsive: 0.0% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of LIME to select highlighted features | Logistic regression · German Credit | % all features responsive: 0.0% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of SHAP to select highlighted features | Logistic regression · German Credit | % all features responsive: 0.0% -> 100.0% | Helped |
| Use Feature Responsiveness scores instead of SHAP actionable to select highlighted features | Logistic regression · Give Me Some Credit | % all features responsive: 22.8% -> 100.0% | Helped |
| Mask immutable features when calculating SHAP scores | Logistic regression explanations · FICO | % at least 1 feature responsive: 22.7% -> 29.7% | Helped |
| Use Feature Responsiveness scores instead of SHAP to select highlighted features | XGBoost · Give Me Some Credit | % all features responsive: 0.0% -> 100.0% | Helped |
Surprising findings
- Action-aware variants of SHAP and LIME still highlight mostly unresponsive features, with only 0.2% of action-aware SHAP explanations fully responsive for logistic regression on HELOC.
- Even among responsive features, about 10.7% of denied individuals can only be approved by decreasing income, so highlighting features alone often fails to convey recourse.
Limitations
- Responsiveness depends on actionability constraints; misspecified constraints can give misleading scores.
- Assumes conservative indisputable constraints, so recourse is not guaranteed since consumers may face additional constraints.
- Experiments use discretized datasets for exact scores; the continuous setting relies on sampling.
Experimental setup
Datasets: HELOC (FICO) Give Me Some Credit German credit
Benchmarks: HELOC German credit Give Me Some Credit
Metrics: % denied % no recourse (fixed prediction) % 1-D recourse % n-D recourse % presented with explanations # features highlighted % all features unresponsive % at least 1 feature responsive % all features responsive
Baselines: SHAP LIME action-aware SHAP action-aware LIME
Builds on: SHAP LIME Agresti-Coull interval
Closest papers by meaning
- Prediction without Preclusion: Recourse Verification with Reachable Sets (ICLR 2024 · similarity 0.89)
- A Market for Accuracy: Classification Under Competition (ICML 2025 · similarity 0.85)
- Fair Classification with Partial Feedback: An Exploration-Based Data Collection Approach (ICML 2024 · similarity 0.84)
- Providing Fair Recourse over Plausible Groups (AAAI 2024 · similarity 0.84)
- SafeAR: Safe Algorithmic Recourse by Risk-Aware Policies (AAAI 2024 · similarity 0.84)
- Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.84)
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.