This study reproduces the effectiveness of MORAL, a post-processing method that improves exposure-based fairness while maintaining competitive utility, and shows that exposure-based metrics uncover biases hidden by $\Delta_\mathrm{DP}$ and that MORAL reduces these biases with minimal utility loss across diverse settings and datasets.
Abstract
In fair ranked link prediction, demographic parity ($\Delta_\mathrm{DP}$) is a common fairness metric. Yet, Mattos et al. (2025) argue that it fails to detect exposure bias because it ignores where links appear in the ranking. In this study, we reproduce this claim by showing that $\Delta_\mathrm{DP}$ can indicate aggregate parity even when some subgroup-pair links are systematically ranked lower than others. The proposed rank-aware Normalized Discounted KL-divergence (NDKL), however, does detect such disparities. We also reproduce the effectiveness of MORAL, a post-processing method that improves exposure-based fairness while maintaining competitive utility. Beyond reproduction, we assess robustness using synthetic homophily settings, categorical sensitive attributes, and additional fairness and utility metrics, including subgroup-pair-adapted Attention-Weighted Rank Fairness (AWRF). Overall, our results show that exposure-based metrics uncover biases hidden by $\Delta_\mathrm{DP}$ and that MORAL reduces these biases with minimal utility loss across diverse settings and datasets. We release a corrected, reproducible implementation at https://github.com/Floris93100/reproducing-MORAL.
Fairness in recommendation systems has drawn growing attention due to rising societal and regulatory concerns over algorithmic bias. Existing fairness-aware approaches typically mitigate bias by either removing sensitive attributes via representation learning or leveraging causal-path interventions (e.g., counterfactual or specific-path debiasing) to distinguish genuine causal effects from confounder-induced correlations between sensitive attributes and user preferences. However, when it comes to evaluation, most prior work adopts both Demographic Parity (DP) and Equal Opportunity (EO) as simultaneous criteria, yet overlooks their inherent tension and the causal nature of the sensitive attribute. Specifically, if a sensitive attribute genuinely drives preference variation, enforcing DP forces equal exposure across groups, contradicting natural interest diversity and severely hurting accuracy; conversely, for spurious correlations, relying solely on EO fails to remove confounder-introduced bias. More importantly, these metrics are typically computed in a static, one-shot manner, ignoring that recommendation is an iterative process where even minor initial disparities can be amplified over time through feedback loops, eventually leading to substantial long-term unfairness. Nevertheless, existing studies rarely address such dynamic, long-term fairness implications, leaving a critical gap in both evaluation and optimization. To resolve this, we propose Long-term Fairness-aware Recommendation via Adaptive Fairness Metric Selection (LFR-via-AFMS). Our framework first learns the causal structure to identify whether the sensitive attribute has a genuine causal effect or merely a spurious association with user preferences. Based on this diagnosis, it adaptively selects the most appropriate fairness criterion: Equal Opportunity for true causality, which allows legitimate group differences in preference, and Demographic Parity for spurious correlations, which eliminates unjustified disparities entirely. The adaptively chosen metric is then integrated into an actor–critic reinforcement learning reward to optimize long-term fairness without sacrificing accuracy. Extensive experiments on Alibaba and MovieLens datasets, with five independent runs and statistical significance testing, demonstrate that the proposed method achieves a superior fairness-accuracy trade-off compared with state-of-the-art baselines, and the adaptive metric selection proves indispensable for maintaining both equity and recommendation quality. We validate the causal diagnosis module through simulation studies with known ground truth and sensitivity analyses confirming robustness across threshold choices.
Algorithmic systems increasingly rank individuals for access to scarce public resources, from child welfare interventions to cancer treatment referrals. The prevailing fairness frame treats disparity as a property of biased data or deficient models, with remedies through calibration and debiasing. Under structural scarcity, where demand exceeds supply by an order of magnitude, allocation becomes a rationing problem, and the statistical properties of ranking diverge sharply from those of classification. We derive a scaling law $D \propto \exp(t \cdot \rho \cdot \Delta)$, in which relative disparity between two groups separated by a structural gap $\Delta$ grows in the product of the scarcity-induced threshold $t$ and rank-discrimination fidelity $\rho$. Scarcity and accuracy interact multiplicatively, producing exponentially larger between-group disparities. We term this dynamic the Accuracy Trap. We validate this Accuracy Trap through Monte Carlo simulation and two independent public-sector systems in Canadian child welfare and U.S. cancer care. Debiasing alone cannot dissolve the trap.
Abstract Background Fairness evaluation is essential for trustworthy clinical risk prediction. However, existing fairness-oriented discrimination metrics either ignore cross-group comparisons or rely on exhaustive pairwise evaluations, making them difficult to interpret and impractical for model selection. Objective This study aimed to develop and evaluate novel fairness-oriented discrimination metrics for clinical risk prediction that address limitations of within-group and pairwise cross-group approaches. Methods We examined theoretical properties of existing U-statistic–based metrics, including concordance index (CI) and area under the receiver operating characteristic curve (AUC), when applied to subgroups. We highlighted the distinction between within-group discrimination (ranking within a subgroup) and group-level discrimination (ranking relative to the broader population). Building on this framework, we proposed group-level extensions of the CI and AUC that summarize subgroup-specific performance in a single interpretable measure. We then applied these metrics to the PREVENT (Predicting Risk of Cardiovascular Disease Events) equation, a recently developed model for atherosclerotic cardiovascular disease. Results The traditional subgroup-specific CI and AUC captured within-group but not group-level discrimination, obscuring inequities in clinical decision-making. Existing cross-group approaches (eg, the xCI and xAUC metrics) addressed this limitation but became computationally and interpretively burdensome with multiple subgroups due to pairwise comparisons. Our proposed metrics provided a streamlined alternative, yielding 1 summary statistic per subgroup while retaining sensitivity to cross-group ranking disparities. Applied to PREVENT, these metrics revealed differences in subgroup performance not apparent from within-group evaluations. Conclusions By distinguishing between within-group and group-level discrimination, our framework clarifies a common source of misinterpretation in fairness evaluation. The proposed group-level extensions of the CI and AUC provide practical, interpretable tools for evaluating fairness in clinical prediction models, enabling more transparent and equitable risk assessment.
Haoyuan Wang, Chuan Hong, Michael J. Pencina et al.· JMIR AI· 0 citations
Fairness audits for LLM-based recommenders have largely focused on observable outputs, implicitly assuming that stable recommendations reflect stable internal processing. We challenge this assumption with FairGap, the first benchmark to jointly evaluate recommendation fairness at two levels: observable output shift (OBS) and hidden representation shift (IBS), measured through controlled counterfactual identity probes across gender, age, and race. Their relationship is summarized via Representation-Output Alignment (ROA), with quadrant diagnostics for identifying user-level hidden-output mismatch. Applied to six open-weight LLM families across three domains, FairGap reveals pervasive hidden-output decoupling: ROA rarely exceeds 0.22, and a non-negligible user population shows stable outputs despite substantial internal shifts, a mode that output-only audits cannot detect by design. Further, activation steering that reduces IBS by up to 8x simultaneously worsens OBS, demonstrating a fundamental tension between internal and output-level fairness that existing frameworks are unequipped to diagnose.
Changyu Lu, Arya Fayyazi, Junhao Zhang et al.· 0 citations
All fairness algorithms unavoidably rely on normative assumptions about fair treatment, yet these assumptions often remain implicit. We argue that these assumptions should be formalized as explicit, auditable policies and introduce FAIR-SHEPHERD, a structural policy-based framework for transparent fairness in real-world settings with noisy labels and shifting distributions. FAIR-SHEPHERD uses gradients as attribution signals, encoded in a Structural Fairness Specification (SFS) that defines vertical coherence and orthogonality to vulnerable proxies. We introduce SFS metrics: Vertical Coherence Score (VCS) to measure directional coherence across adjacent normative slices, and Horizontal Leakage Score (HLS), augmented by a signed directional variant, to detect gradient alignment with policy-declared vulnerable or proxy directions. To enforce these policies, we develop S-agnostic tools that optimize policy-aligned gradient objectives using lattice-defined proxies. We demonstrate that outcome-based auditing is brittle to measurement error: under 10% label noise on Adult, Worst-Group AUC for ERM drops by 0.159. In contrast, our gradient-based structural metrics provide a label-agnostic audit of the decision logic, remaining stable even when evaluation labels are corrupted. Specialized fairness baselines including ARL, JTT, and GoG retain substantial structural leakage on COMPAS, with HLS values from about 0.47 to 0.71. Some also reduce EOD relative to ERM, which shows that outcome and structural criteria can diverge. Using our S-agnostic Gradient Penalty tools, we reduce policy-specified structural leakage by roughly 85% on COMPAS (0.478 → 0.073) and over 94% on Adult (0.138 → 0.008) while retaining competitive AUC and Worst-Group AUC. We have released our source code to facilitate further research here.
Shubham Mohole, Hongjun Choi, Christian Schroeder et al.· ACM Journal on Responsible C...· 0 citations
Graph Counterfactual Fairness (GCFair), a novel framework that achieves counterfactual fairness by explicitly identifying and disentangling the subsets of node features and graph structures genuinely affected by sensitive attributes, is proposed.
Zichong Wang, Zhipeng Yin, Zhong Chen et al.· 3 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
