Aug 2026· Journal of Artificial Intelligence Research· 0 citations
TL;DR
This work introduces a method to lower-bound the discrepancy of a classifier: a quantity that jointly captures inaccuracy and unfairness, and develops a computationally efficient procedure for calculating the tightest possible lower bound on the classifier’s discrepancy.
Abstract
Background: Classifiers increasingly affect people’s lives, necessitating their audit for fairness and accuracy on diverse populations. However, direct auditing is often not possible, due to a lack of access to the classifier or to suitable individual-level validation data.
Objectives: This work aims to assess the fairness and accuracy of black-box classifiers using only population-level statistics, without requiring access to the classifier or individual predictions. Specifically, it introduces a method to lower-bound the discrepancy of a classifier: a quantity that jointly captures inaccuracy and unfairness.
Methods: We define a novel measure of unfairness based on the equalized odds fairness criterion, quantifying the fraction of the population on which a classifier deviates from ideal fair behavior. Using this measure, we develop a computationally efficient procedure for calculating the tightest possible lower bound on the classifier’s discrepancy, using only aggregated rates of positive predictions and true positives across protected sub-populations.
Results: Empirical evaluations confirm the tightness of the proposed lower bound in practical settings. The method is demonstrated on several use cases, including estimating the reliability of voting polls and assessing the fairness of patient identification from internet search data. The code and data are available at https://github.com/sivansabato/bfa2.
Conclusions: This work provides a practical and interpretable framework for auditing classifiers using population-level statistics. The proposed approach enables stakeholders to identify fairness and accuracy concerns in settings where traditional auditing is not feasible.
For a fair society, decisions impacting people's lives must be taken fairly, irrespective of their protected characteristics. To achieve algorithmic fairness, appropriate measures must guide machine learning methods toward fair(er) decision recommendations. There are numerous fairness metrics for classification; however, for regression the literature is still developing. Most fair machine learning methods combine fairness and error in a single objective. Instead, we apply multi-objective optimization, allowing the fairness metrics to be optimized alongside the error. We develop a transition from classification to regression for symbolic regression by first discretizing the task into a fixed number of classes, progressively increasing their number, and finally by using full regression fairness metrics. We demonstrate the framework on the Law School Admission Council dataset. While fairness objectives do not appear to be significantly different among methods, accuracy is clearly better when applying full regression fairness.
Michel El Saliby, David Ibehej, Nicole Vergara et al.· Proceedings of the Genetic a...· 0 citations
With the continuous growth in data volumes, efficient data selection has become critical for scaling machine learning (ML) classifiers. Although the existing methods aim to accelerate training without compromising accuracy, their impact on fairness and broader utility remains largely unexplored. This paper presents the first quantitative and qualitative evaluation of these effects through a large-scale empirical analysis comprising 24,395 experiments across 9 real-world datasets, 7 ML models, and several state-of-the-art data selection methods, conducted using ELAPSE, our open-source evaluation framework. Our analysis reveals several key findings and non-trivial observations. The results show that ML data selection can hurt model fairness in a non-negligible number of cases, and compromise model utility in more than half of the cases. These effects differ across datasets, with no single method consistently outperforming the others, and in nearly one-third of the cases, either utility or fairness must be prioritized. Surprisingly, combining data selection with classical bias mitigation increases model unfairness and often degrades model utility. These findings provide interesting research directions for utility- and fairness-aware ML data selection.
Nawel Benarba, Zeyang Kong, Sara Bouchenak· IEEE Access· 1 citation
When a classifier is trained by actively choosing which examples to label, should its labeling budget depend on which notion of fairness it must satisfy? We answer this with the first information-theoretic lower bounds for group-fair active learning, and the answer is a sharp hierarchy across the three standard notions. Demographic Parity (DP), which equalizes the positive prediction rate across groups, can be verified from unlabeled data alone, so it adds no labels beyond ordinary learning. Equal Opportunity (EO), which equalizes each group's true positive rate, is fundamentally harder: estimating a rate conditioned on the true label forces the learner to collect labeled positives from every group, an irreducible cost that no query strategy can avoid. Equalized Odds (EqOdds), which additionally equalizes false positive rates, costs no more than EO whenever positives are the minority, the typical regime. This exposes a separation at the heart of fair learning: active querying yields the well-known exponential savings for accuracy, but no speedup whatsoever for fairness verification. Concretely, certifying EO to fairness tolerance γ requires Ω(d/ε2 + k/(ε2 p+)) labels, where d is the model's Vapnik--Chervonenkis (VC) dimension, k the number of groups, ε the target error, and p+ = Pr[Y=1] the positive rate; the fairness term is independent of ε, so it is a floor that active learning cannot lower. We give matching algorithms (ConstrainedERM-DP for DP, FairStratified for EO and EqOdds) that are tight up to logarithmic factors. The cost is largest exactly when positives are rare: at p+ = 0.01, as in fraud detection, EO demands roughly 90× more labels than DP. Experiments on Folktables, COMPAS, and Adult confirm the predicted scaling and yield 36 to 39% label savings over passive fair-learning baselines.
Chang Lu, Yizheng Zhao· Proceedings of the 32nd ACM...· 0 citations
: This paper proposes a robust, reliable, and reproducible methodology for evaluating group fairness in classification algorithms. Building upon established theoretical definitions of non-discrimination criteria - Independence, Separation, and Sufficiency-we present a comprehensive approach to quantifying fairness. Notably, our methodology extends beyond binary group comparisons to accommodate scenarios with multiple sensitive groups. Furthermore, we provide a methodological approach for assessing fairness within the Sufficiency criterion by operationalizing Calibration, elucidating critical issues and conceptual subtleties that appear to have been overlooked in existing literature. We assess the reliability and robustness of our model through applications to one real dataset.
Massimiliano Mancini, D. Merlini, M. Verri· Proceedings of the 15th Inte...· 0 citations
Results indicate that global feature importance, used as an active search signal rather than a post-hoc diagnostic, improves both the effectiveness and the efficiency of individual fairness testing.
H. Mamman, Abdullateef Oluwagbemiga Balogun, Mustapha Maidawa et al.· Journal of King Saud Univers...· 0 citations
Algorithmic fairness refers to the equitable treatment of users regardless of their identity, demographics, or other personal characteristics, whereas biased AI may produce unfair outcomes when the input data reflect imbalances or underrepresentation of certain groups. This systematic literature review analyzes studies published between 2018 and 2025 that report approaches for handling bias in data and datasets, with particular attention to dataset-centered techniques. We examine how bias is identified, mitigated, and reflected in model outcomes. Our method defines research questions and details the search strategy, study selection procedure, and protocols for data extraction and analysis. The review maps fairness dimensions, the metrics employed, pre-processing and in-processing methods, the role of human participation, observed effects and side effects, and the ethical considerations that inform practice. The evidence suggests that proactive, iterative approaches incorporating human oversight are essential for mitigating algorithmic discrimination and enhancing the trustworthiness of AI systems. Neglecting these considerations risks significant societal harm and the perpetuation of bias against marginalized groups in high-stakes contexts, particularly in an era where AI is increasingly widespread.
Matheus de Morais Leca, M. Bento, Ronnie de Souza Santos· SIGSOFT FSE Companion· 0 citations
