Research map / Learning theory and algorithms / Prediction and decision losses
Learning Models for Actionable Recourse
Alexis Ross, Himabindu Lakkaraju, Osbert Bastani
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Train with the recourse loss (λ=0.8) instead of the baseline loss | Neural network classifier · Adult, Compas, Bail, German | recourse neg using adversarial training: 0.007 -> 0.498; 0.000 -> 0.967; 0.000 -> 0.993; 0.127 -> 0.968 | Improved |
| Train with the recourse loss (λ=0.8) instead of the baseline loss | Neural network classifier · Adult, Compas, Bail, German | F1 score: 0.697 -> 0.636; 0.739 -> 0.717; 0.775 -> 0.760; 0.447 -> 0.419 | Hurt |
| Train with the recourse loss (λ=0.8) instead of the baseline loss | Neural network classifier · Adult, Compas, Bail, German | recourse neg using gradient descent: 0.210 -> 0.496; 0.579 -> 1.000; 0.317 -> 1.000; 0.804 -> 0.864 | Improved |
| Train with the recourse loss (λ=0.8) instead of the baseline loss | Neural network classifier · Adult, Compas, Bail, German | recourse neg using linear approximation: 0.220 -> 0.053; 0.156 -> 0.068; 0.102 -> 0.029; 0.204 -> 0.086 | Hurt |
| Use the adversarial-training recourse solver rather than gradient descent | Models trained with the recourse loss · Adult, Compas, Bail, German | recourse neg: 0.498 vs 0.496; 0.967 vs 1.000; 0.993 vs 1.000; 0.968 vs 0.864 | Mixed |
| Choose the threshold to satisfy PARE instead of maximizing F1 | Baseline classifier · Adult | F1 score: 0.697 -> 0.400 | Hurt |
| Choose the threshold to satisfy PARE instead of maximizing F1 | Recourse-trained classifier · Adult | F1 score: 0.636 -> 0.526 | Hurt |
| Choose the threshold to satisfy PARE instead of maximizing F1 | Recourse-trained classifier · Compas, Bail, German | F1 score: 0.717 -> 0.721; 0.760 -> 0.776; 0.419 -> 0.457 | No clear effect |
| Vary recourse-loss weight λ from zero to positive values (ablation) | Neural network classifier · Adult, Compas, Bail, German | recourse rate | Improved |
| Use linear approximation instead of gradient descent or adversarial training to compute recourse (ablation) | Recourse-trained classifier · Adult, Compas, Bail, German | recourse rate: 0.053 vs 0.496; 0.068 vs 1.000; 0.029 vs 1.000; 0.086 vs 0.864 | Hurt |
| Use recourse loss (λ=0.25) instead of baseline loss (λ=0) | Sentiment classifier · Stanford Sentiment Treebank test set | recourse neg: 0.239 -> 0.950 | Improved |
| Use recourse loss (λ=0.25) instead of baseline loss (λ=0) | Sentiment classifier · Stanford Sentiment Treebank test set | F1 score: 0.881 -> 0.864 | Hurt |
Closest papers by meaning
- Prediction without Preclusion: Recourse Verification with Reachable Sets (ICLR 2024 · similarity 0.88)
- Providing Fair Recourse over Plausible Groups (AAAI 2024 · similarity 0.86)
- Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse (ICLR 2025 · similarity 0.85)
- SafeAR: Safe Algorithmic Recourse by Risk-Aware Policies (AAAI 2024 · similarity 0.85)
- ElliCE: Efficient and Provably Robust Algorithmic Recourse via the Rashomon Sets (NeurIPS 2025 · similarity 0.85)
- From Search to Sampling: Generative Models for Robust Algorithmic Recourse (ICLR 2025 · 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.
