Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Referee-Meta-Learning for Fast Adaptation of Locational Fairness
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 Meta-Ref, a locational meta-referee that learns per-location learning rates within a MAML-style meta-learner so that predictions are fairer across locations in unseen test regions, evaluated on crop classification and traffic accident risk estimation.
Problem
Machine learning models show implicit locational biases (prediction quality disparity across locations), and existing fairness methods do not handle fairness when the evaluated locations differ between training and test regions, especially with few labeled samples.
Approach
A locational meta-referee network assigns location-dependent learning rates to mini-batches from different locations based on their relative performance and encodings, trained with a three-phase MAML-style scheme that includes a fairness-loss meta-update for both the referee and the predictor; the referee is frozen during test-region fine-tuning.
Key results
- On crop classification task set 1, Meta-Ref reduced LF to 0.107 and ALF to 0.107 versus MAML's 0.119 and 0.119, with F1 of 0.716 vs 0.719.
- On traffic accident estimation task set 2, Meta-Ref achieved LF 0.081 and ALF 0.083 versus MAML's 0.084 and 0.087, with RMSE 0.136 vs 0.133.
- Ablation on crop classification shows removing the performance-to-predictor gradient (MR-P2P) drops F1 to 0.606 and raises LF to 0.176.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Meta-Ref instead of MAML | DNN for crop classification · Crop classification, task sets 1–3 | LF: 0.119 -> 0.107; 0.124 -> 0.114; 0.126 -> 0.118 | Helped |
| Use Meta-Ref instead of MAML | DNN for crop classification · Crop classification, task sets 1–3 | F1: 0.719 -> 0.716; 0.718 -> 0.710; 0.714 -> 0.706 | Mixed |
| Use Meta-Ref instead of MAML | LSTM for traffic risk estimation · Traffic risk estimation, task sets 1–3 | LF: 0.108 -> 0.102; 0.084 -> 0.081; 0.087 -> 0.085 | Helped |
| Use Meta-Ref instead of MAML | LSTM for traffic risk estimation · Traffic risk estimation, task sets 1–3 | RMSE: 0.134 -> 0.131; 0.133 -> 0.136; 0.137 -> 0.138 | Mixed |
| Remove the performance gradient update to the prediction model (P2P) (ablation) | Meta-Ref on crop classification · Crop classification, ablation study | LF: 0.112 -> 0.176 | Hurt |
| Remove the fairness gradient update to Meta-Ref (F2M) (ablation) | Meta-Ref on crop classification · Crop classification, ablation study | LF: 0.112 -> 0.143 | Hurt |
| Remove the fairness gradient update to the prediction model (F2P) (ablation) | Meta-Ref on crop classification · Crop classification, ablation study | LF: 0.112 -> 0.131 | Hurt |
| Set fairness loss scaling factor λ to values from 0.01 to 10 instead of 0 (ablation) | Meta-Ref on crop classification · Crop classification sensitivity analysis | LF: 0.122 -> 0.115; 0.122 -> 0.112; 0.122 -> 0.115; 0.122 -> 0.114 | Helped |
Surprising findings
- Domain-adaptation variants often do not improve fairness over plain baselines and sometimes worsen it.
- On traffic accident estimation, Meta-Ref's fairness gain over MAML is small and in some task sets (e.g., task set 3 RMSE) it is slightly worse on prediction.
Limitations
- Evaluated on only two case studies with fixed spatial setups (one crop region, one city-grid dataset).
- Fairness is defined as prediction quality parity via standard deviation of per-location scores, not other fairness notions.
- Ablation and sensitivity results are reported on a single task set for crop classification.
Experimental setup
Datasets: Sentinel-2 imagery with USDA Crop Data Layer labels (Central Valley, California) Iowa traffic accident records (3 years, 47 factors)
Benchmarks: Sentinel-2 crop classification (walnut mapping) Iowa traffic accident risk estimation
Metrics: F1 RMSE LF (locational fairness, std of per-location performance) ALF (adjusted locational fairness)
Baselines: DNN LSTM Reg Adv Dom-DNN Dom-LSTM Dom-Reg Bi-Lvl MAML
Builds on: MAML Bi-level fairness adjustment domain adaptation
Closest papers by meaning
- Task-Adaptive Meta-Learning Framework for Advancing Spatial Generalizability (AAAI 2023 · similarity 0.85)
- Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting (AAAI 2025 · similarity 0.82)
- Fairness-Aware Meta-Learning via Nash Bargaining (NeurIPS 2024 · similarity 0.82)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.80)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.80)
- Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales (NeurIPS 2026 · similarity 0.80)
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.