Research map / Reinforcement learning / Policy gradient and actor-critic
SafeAR: Safe Algorithmic Recourse by Risk-Aware Policies
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 Safe Algorithmic Recourse (SafeAR), computing risk-aware recourse policies via a greedy risk-sensitive value iteration (G-RSVI) on finite-horizon MDPs, evaluated with VaR/CVaR on Adult Income and German Credit.
Problem
Algorithmic recourse methods optimize feasibility, sparsity and proximity of sequential actions but ignore the risk of higher-than-average costs arising from uncertain action outcomes.
Approach
Formulate recourse as a finite-horizon MDP and compute policies maximizing mean minus beta times standard deviation of total cost via a greedy backward value iteration (G-RSVI), letting individuals choose a risk-aversion level; apply VaR and CVaR from finance to summarize risk.
Key results
- On Adult Income with horizon 12, beta=0.5 reduces cost variance from 1.23 to 0.77 and VaR95 from 4.76 to 4.44 relative to risk-neutral beta=0, at a mean cost increase from 3.49 to 3.59.
- On German Credit example instance, beta=0.75 lowers cost variance to 0.65 versus 3.49 at beta=0.
- Risk-averse policies show a gender disparity in risk measures, with the female group exposed to higher risk in Adult Income, and the gap widens with beta in that dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a cost-variance penalty (β=0.5) instead of using the risk-neutral policy | G-RSVI on Adult Income, H=12 · Adult Income (n=2000) | cost variance: 1.26 -> 0.83 | Helped |
| Add a cost-variance penalty (β=0.5) instead of using the risk-neutral policy | G-RSVI on German Credit, H=12 · German Credit (n=281) | cost variance: 0.48 -> 0.30 | Helped |
| Add a cost-variance penalty (β=0.5) instead of using the risk-neutral policy | G-RSVI on Adult Income, H=12 · Adult Income (n=2000) | mean cost: 3.69 -> 3.78 | Hurt |
| Add a cost-variance penalty (β=0.5) instead of using the risk-neutral policy | G-RSVI on German Credit, H=12 · German Credit (n=281) | mean cost: 1.65 -> 1.70 | Hurt |
| Increase risk-aversion to β=2.0 (ablation) | G-RSVI on Adult Income, H=12 · Adult Income (n=2000) | success rate: 0.995 -> 0.717 | Hurt |
| Increase risk-aversion to β=2.0 (ablation) | G-RSVI on Adult Income, H=12 · Adult Income (n=2000) | cost variance: 1.26 -> 0.28 | Helped |
| Increase risk-aversion to β=2.0 (ablation) | G-RSVI on German Credit, H=12 · German Credit (n=281) | cost variance: 0.48 -> 0.65 | Hurt |
| Restrict favorable goal states to the 10 DiCE counterfactuals (ablation) | G-RSVI on Adult Income · Adult Income (n=100), age <30 | success rate: 0.997 -> 0.774 | Hurt |
| Restrict favorable goal states to the 10 DiCE counterfactuals (ablation) | G-RSVI on German Credit · German Credit (n=100) | success rate: 1.000 -> 0.980 | Hurt |
| Use Lower Partial Standard Deviation instead of standard deviation as the risk penalty (ablation) | G-RSVI on Adult Income and German Credit · Adult Income and German Credit | risk measures | No clear effect |
| Increase the horizon from H=4 to H=12 (ablation) | G-RSVI on Adult Income · Adult Income | success rate: 0.335 -> 0.717 | Helped |
| Increase risk-aversion from β=0 to β=0.5 (ablation) | G-RSVI on German Credit, H=4 · German Credit | success rate: 0.963 -> 0.972 | Helped |
Surprising findings
- Policies with better sparsity and proximity do not correspond to lower-risk policies.
- Restricting recourse to DiCE counterfactual goal states lowers success rate and yields worse risk measures than the full-state-space MDP policy.
Limitations
- G-RSVI is a greedy heuristic and does not guarantee the optimal risk-sensitive policy.
- Action costs and transition probabilities are set heuristically by domain knowledge rather than learned from data.
- Limited to discrete states, discrete actions, and finite horizon MDPs.
Experimental setup
Datasets: UCI Adult Income German Credit Health Insurance Premiums (Kaggle)
Benchmarks: UCI Adult Income German Credit
Metrics: success rate mean cost cost variance VaR CVaR sparsity proximity
Baselines: FASTAR DiCE10 risk-neutral policy (beta=0)
Builds on: FASTAR DiCE risk-sensitive reinforcement learning
Closest papers by meaning
- Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents (ICML 2025 · similarity 0.85)
- Providing Fair Recourse over Plausible Groups (AAAI 2024 · similarity 0.85)
- Algorithmic Recourse for Long-Term Improvement (ICML 2025 · similarity 0.85)
- Pessimism Meets Risk: Risk-Sensitive Offline Reinforcement Learning (ICML 2024 · similarity 0.85)
- Risk-Conditioned Reinforcement Learning: A Generalized Approach for Adapting to Varying Risk Measures (AAAI 2024 · similarity 0.84)
- Towards Safe Policy Learning under Partial Identifiability: 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.