Algorithmic Fairness
A Unified Formulation for Multiclass and Regression Targets under Postprocessing-Based Guarantees
Generated automatically from the limitations stated in 10 papers (ICML, ICLR, CVPR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Ten independent lines of work — spanning fairness postprocessing, conformal prediction, and related certification-style methods — each stop at binary or, at best, categorical classification targets. Continuous outcomes and structured prediction are simply out of reach with current formulations, which means a large fraction of real prediction problems (prices, dosages, rankings, likelihoods) get either no guarantee at all or a crude discretization that distorts the guarantee. Practitioners must choose between losing the guarantee and losing the task, and every paper defers the same extension to "future work" rather than solving it jointly.
Why it matters
Every one of the ten methods becomes applicable to regression and multiclass problems without redesigning each from scratch, and the theory gives a single error-accounting template that future postprocessing results can inherit instead of re-deriving.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Re-derive the core postprocessing condition (Bayes-optimality / exchangeability / group-conditional calibration) for a continuous target, starting from a discretization-and-refinement argument: bin the target, apply the existing binary machinery per bin, then characterize the approximation error as bin width shrinks. Measure: coverage / fairness-gap of the resulting predictor versus a binning-free oracle, as a function of resolution.
- 2
Extend the multiclass case first, since it is the smaller step: replace the single threshold with a score-vector construction, prove the guarantee generalizes from binary to K classes with a factor depending on K, and measure the empirical gap between the K-class guarantee and K one-vs-rest binary guarantees on standard tabular benchmarks.
- 3
Test on regression by mapping the continuous problem to a binary one on derived events (e.g., exceedance probabilities across a grid of thresholds), verifying that the family of per-threshold guarantees stitches into a valid guarantee over the full output space; measure integrated coverage/calibration error against conformalized baselines.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If the family of binary guarantees cannot be shown to jointly hold without strong independence or union-bound assumptions, the stitched regression/multiclass guarantee becomes vacuous or too loose to be useful empirically.
Sub-problems
Benchmarking and Extending Conformal Uncertainty Quantification Across Regression and Structured Prediction Tasks
Existing uncertainty quantification and conformal prediction methods are predominantly formulated and evaluated solely on classification benchmarks, leaving their empirical behavior on continuous and structured outputs uncharacterized. Practitioners in continuous domains (such as regression or object detection) cannot reliably determine whether non-conformity scores and theoretical coverage guarantees transfer or degrade under real-world continuous and spatial distributions. Without cross-task evaluation, it remains unknown whether these methods suffer systematic failure modes when transitioning away from discrete label spaces.
Evidence
Each paper's own statement of the limitation, verbatim.
- Semi-Supervised Conformal Prediction With Unlabeled Nonconformity ScoreCVPR 2026
Designed and evaluated exclusively for classification tasks and does not extend to regression settings.
- Rank-Guided Pseudo-Bias Learning for Robust Black-Box AdaptationCVPR 2026
Limited to classification tasks, with no support for structured outputs or regression.
- RECAST: Model Reconstruction via Counterfactual-Aware Wasserstein Geometry under Limited DataICML 2026
Restricted to binary classification settings and does not currently extend to multi-class target models.
- MultiFPT: Towards Multi-Attribute Fairness in Pre-Trained Graph Neural Networks via Prompt TuningTheWebConference 2026
Evaluated primarily under binary sensitive attributes and binary node classification tasks.
- Measuring Uncertainty CalibrationICLR 2026
Formulated strictly for binary classification tasks, with multiclass extensions left as future work.
- MCGrad: Multicalibration at Web ScaleKDD 2026
Currently restricted to binary classification tasks, lacking extensions to multiclass classification or continuous regression.
- Fair Graph Machine Learning under Adversarial Missingness ProcessesICLR 2026
Empirically validated and formulated only for binary classification tasks and binary sensitive attributes.
- Fair Dataset Distillation via Cross-Group Barycenter AlignmentICML 2026
Restricted to image classification tasks under single discrete protected attributes.
Show all 10 papers
- Courtroom Analogy: New Perspective on Uncertainty-Aware ClassificationICML 2026
Restricted to classification settings and does not currently support continuous regression tasks.
- Unprocessing Seven Years of Algorithmic FairnessICLR 2024
Restricted to tabular data; high-dimensional raw-feature datasets (images, text) are not explored, where the Bayes-optimality assumption justifying postprocessing may fail.
Nearest existing work
- Fair and Optimal Classification via Post-ProcessingICML 2023
- Fair Regression: Quantitative Definitions and Reduction-Based AlgorithmsICML 2019
- Fair Wrapping for Black-box PredictionsNeurIPS 2022
- Beyond Procedure: Substantive Fairness in Conformal PredictionICML 2026
- Regression under demographic parity constraints via unlabeled post-processingNeurIPS 2024
- Pushing the limits of fairness impossibility: Who's the fairest of them all?NeurIPS 2022
- Equal Opportunity of Coverage in Fair RegressionNeurIPS 2023
- Unprocessing Seven Years of Algorithmic FairnessICLR 2024
- A Generic Framework for Conformal FairnessICLR 2025
- Fair Conformal Classification via Learning Representation-Based GroupsICLR 2026
- Conformalized Fairness via Quantile RegressionNeurIPS 2022
- Beyond Adult and COMPAS: Fair Multi-Class Prediction via Information ProjectionNeurIPS 2022
- FRAPPÉ: A Group Fairness Framework for Post-Processing EverythingICML 2024
- Demystifying the Optimal Fair Classifier in Multi-Class ClassificationICML 2026
- Fair Generalized Linear Models with a Convex PenaltyICML 2022
Related open problems
Algorithmic Fairness
Benchmarking Algorithmic Fairness Methods Under Violations of Sensitive Attribute Preconditions
Existing algorithmic fairness methods are almost universally evaluated under rigid preconditions: fully observed or missing-at-random discrete group labels, bounded minority group representations ($\rho$-fraction), and absence of feature-dependent annotation noise. In real-world deployments, sensitive attributes are often unannotated, continuous, severely underrepresented, or corrupted by feature-dependent noise. Because fairness methods have not been systematically evaluated across these non-idealized regimes, practitioners cannot determine which, if any, algorithmic debiasing approaches retain their fairness guarantees or utility tradeoffs when theoretical preconditions fail.
Algorithmic Fairness
Robustness and Sensitivity of Fairness Interventions Under Misspecified Merit Distributions and Latent Parameters
Many algorithmic fairness frameworks rely on theoretical assumptions that cannot be verified in deployment, specifically requiring known ground-truth merit distributions, unobservable transfer constants, or rigid pre-specified group partitions. In real-world applications, practitioners have access only to observed proxy labels and imperfect covariates, leaving the true merit distributions and transfer constants unknown and unbounded. Consequently, when fairness interventions are calibrated against estimated or misspecified latent parameters, theoretical parity guarantees can break down unpredictably, risking both utility loss and unintended disparity.