Research map / Optimization / Optimizers
From Search to Sampling: Generative Models for Robust Algorithmic 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 GenRe, a transformer encoder-decoder that learns a conditional recourse likelihood from synthetically paired negative/positive instances and produces recourse by forward sampling rather than gradient search.
Problem
Algorithmic recourse must jointly balance proximity, plausibility, and validity, but existing methods train these objectives separately and rely on non-robust gradient-based search at inference; there is also no direct recourse supervision for training generative recourse models.
Approach
Train an autoregressive conditional generative model R(x+|x) on pairs sampled from positive-class data with importance weights exp(-lambda*cost), so that the three recourse criteria are learned jointly; recourse is then obtained by forward sampling, with a consistency result for the pair-sampling estimator.
Key results
- GenRe achieves the best combined Score on all three datasets: 1.93 (Adult Income), 1.89 (COMPAS), 1.90 (HELOC), with validity 1.00/0.99/1.00 and LOF 0.98/0.97/1.00.
- On Adult Income, GenRe's cost is 0.69 versus CRUDS 3.17 and TAP 1.10, with comparable validity and higher LOF.
- Among baselines, CRUDS is second best (Score 1.72/1.83/1.37), while ROAR and PROBE yield very high cost and low LOF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train a joint conditional recourse likelihood model and sample recourse forward | GenRe · Adult Income | Score: 1.72 -> 1.93 | Helped |
| Train a joint conditional recourse likelihood model and sample recourse forward | GenRe · COMPAS | Score: 1.83 -> 1.89 | Helped |
| Train a joint conditional recourse likelihood model and sample recourse forward | GenRe · HELOC | Score: 1.37 -> 1.90 | Helped |
| Add classifier-confidence filtering (gamma=0.7) to nearest-neighbor recourse | NNR · Adult Income, COMPAS, HELOC | Score: 1.34 -> 1.53 (Adult Income); 1.71 -> 1.90 (COMPAS); 1.53 -> 1.62 (HELOC) | Helped |
| Also require observed training label y=1 for nearest-neighbor recourse | NNR with gamma=0.7 · Adult Income, HELOC | Score: 1.53 -> 1.71 (Adult Income); 1.62 -> 1.88 (HELOC) | Helped |
| Use conditional recourse likelihood instead of unconditional positive-class density (ablation) | GenRe · Synthetic datasets | Recourse density alignment | Helped |
| Vary the inference temperature and sampling sigma (ablation) | GenRe · Adult Income, COMPAS, HELOC | Score: Adult Income: 1.83–1.94; COMPAS: 1.67–1.90; HELOC: 1.87–1.91 | Mixed |
| Increase the balance parameter lambda (ablation) | GenRe · Adult Income | Score: 1.8 (lambda=0.5); 1.78 (lambda=1.0); 1.83 (lambda=2.5); 1.93 (lambda=5.0); 1.92 (lambda=10.0) | Mixed |
| Train with less data (ablation) | GenRe · Adult Income | Score: 1.93 (1.0 fraction); 1.69 (0.1); 1.90 (0.2); 1.82 (0.4); 1.85 (0.8) | Mixed |
| Use TabSyn with SVDD guidance instead of paired GenRe training | TabSyn · Adult Income, COMPAS, HELOC | Score: 1.93 vs 1.58 (Adult Income); 1.89 vs 1.37 (COMPAS); 1.90 vs 1.80 (HELOC) | Hurt |
Surprising findings
- Nearest-neighbor recourse restricted to positive-labeled points (NNR with gamma=0.7, y=1) achieves a Score of 1.90 on COMPAS, matching GenRe's 1.89 there.
- Pretrained TabSyn diffusion with SVDD guidance performs worse than GenRe, often failing immutability constraints.
Limitations
- Results rely on a Random Forest labeler as the gold classifier, so validity is measured against a simulated latent model rather than real decisions.
- Performance on COMPAS is comparable to NNR variants, so the advantage is not uniform across datasets.
- Experiments are limited to three small tabular datasets with binary-encoded categoricals.
Experimental setup
Datasets: Adult Income COMPAS HELOC
Benchmarks: Adult Income COMPAS HELOC
Metrics: cost (l1) validity LOF inlier fraction combined score
Baselines: Wachter GS DICE ROAR PROBE TAP CCHVAE CRUDS NNR TabSyn+SVDD
Builds on: CRUDS CCHVAE REVISE TabSyn SVDD
Closest papers by meaning
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.86)
- ElliCE: Efficient and Provably Robust Algorithmic Recourse via the Rashomon Sets (NeurIPS 2025 · similarity 0.85)
- XTC: Head-Aware Sampling by Excluding Top Choices (NeurIPS 2026 · similarity 0.85)
- Providing Fair Recourse over Plausible Groups (AAAI 2024 · similarity 0.85)
- Revolve: Optimizing AI Systems by Tracking Response Evolution in Textual Optimization (ICML 2025 · similarity 0.85)
- A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops (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.