Online Primal-Dual Allocation Prevents Fairness Collapse in Tabular SSL
A recent study published on arXiv (2605.16446) reveals two failure modes in semi-supervised learning for tabular data when fairness constraints are integrated with confidence-gated pseudo-labeling. The first mode, Masking Collapse, occurs when fairness regularization diminishes model confidence, leading to a lack of pseudo-labels. The second, Trivial Saturation, results in the model converging to constant predictors. To tackle these issues, the authors introduce Online Primal-Dual Allocation (OPDA), an online controller that adjusts fairness and entropy-based stability penalties in real-time, utilizing signals from violation, risk, and pseudo-label health. Tested on Adult, ACSIncome, and COMPAS datasets, OPDA effectively addresses the problems seen with static weighting and basic adaptive methods, particularly in critical areas like medical diagnosis and credit scoring.
Key facts
- Paper arXiv:2605.16446 identifies two failure modes in tabular fair SSL: Masking Collapse and Trivial Saturation.
- Masking Collapse occurs when fairness regularization erodes confidence, starving pseudo-labels.
- Trivial Saturation involves drift to constant predictors under confidence-gated pseudo-labeling.
- Online Primal-Dual Allocation (OPDA) is proposed as an online controller for fairness and entropy penalties.
- OPDA uses violation, risk, and pseudo-label health signals to schedule penalties dynamically.
- Evaluated on Adult, ACSIncome, and COMPAS tabular benchmarks.
- OPDA mitigates degenerate regimes compared to static weighting and simple adaptive baselines.
- The work targets high-stakes applications: medical, credit, and recidivism prediction.
Entities
Institutions
- arXiv