Avoiding Structural Failure Modes in Tabular Fair SSL: Online Primal-Dual Allocation under Confidence Gating

Researchers identify critical failure modes in semi-supervised learning (SSL) applied to tabular data with fairness constraints, where fairness regularizers can paradoxically erode model performance. They propose Online Primal-Dual Allocation (OPDA), an adaptive controller that dynamically balances fairness and stability penalties without manual tuning, demonstrating improved robustness across benchmark datasets like Adult, COMPAS, and ACSIncome.

This research addresses a fundamental tension in high-stakes machine learning: applying fairness constraints to semi-supervised systems on tabular data (medical records, credit decisions, recidivism assessments) can create unexpected pathological behaviors. The researchers identified two specific failure modes through systematic stress testing: Masking Collapse, where fairness penalties reduce model confidence and starve the pseudo-labeling pipeline, and Trivial Saturation, where models degenerate into constant predictors. These phenomena represent a structural conflict rather than a tuning problem.

The proposed OPDA controller represents a methodological advance by framing fairness-utility trade-offs as an online optimization problem rather than a static hyperparameter selection task. Instead of manually choosing a single fairness weight per dataset, OPDA uses three signal types—violation metrics, risk scores, and pseudo-label health indicators—to dynamically adjust penalty weights during training. This approach eliminates the need for expensive cross-validation or per-dataset calibration.

The implications extend across regulated industries where algorithmic fairness compliance is mandatory. Current practices typically rely on practitioners selecting fixed fairness weights empirically, a process that scales poorly and often produces unstable results. OPDA's calibration-free design could substantially reduce implementation friction for deploying fair SSL systems in production environments. The evaluation across three canonical fairness benchmarks demonstrates that the controller reaches competitive performance points without domain-specific tuning, though trade-offs between fairness and utility remain architecture-dependent.

Future work should test OPDA's generalization to other SSL paradigms, higher-dimensional data, and real-world deployment scenarios where feedback loops and distribution shift occur continuously.

• →Fairness regularizers in semi-supervised tabular learning can trigger two failure modes: Masking Collapse and Trivial Saturation, revealing fundamental conflicts in the approach.
• →OPDA eliminates manual fairness weight tuning by dynamically scheduling penalties using violation, risk, and pseudo-label health signals.
• →The method achieves non-degenerate operating points across Adult, COMPAS, and ACSIncome benchmarks without per-dataset calibration.
• →Fairness-utility trade-offs remain dataset-dependent, suggesting no universal solution exists but OPDA provides better calibration-free defaults.
• →This addresses a practical bottleneck for deploying fair SSL in regulated industries like credit, healthcare, and criminal justice systems.