This work identifies routing-induced bias, a failure mode in which subgroup imbalance drives the gating network to route subgroups onto a few experts, and proposes an end-to-end Mixture-of-Experts (MoE) framework that corrects it and improves fairness while maintaining competitive predictive performance.
Abstract
Deep learning models often produce performance disparities across demographic groups, due to the training data imbalance with respect to sensitive attributes such as gender or age. To address this problem, existing work has explored fair representation learning, data re-sampling, and adversarial training, which can be broadly categorized into two main approaches. Single-stage methods typically learn a shared representation for fairness, but often struggle to handle heterogeneous subgroup distributions. Two-stage methods learn representations separately from the final prediction task, which can lead to misalignment between fairness objectives and downstream predictions. We identify routing-induced bias, a failure mode in which subgroup imbalance drives the gating network to route subgroups onto a few experts, and propose an end-to-end Mixture-of-Experts (MoE) framework that corrects it. Specifically, we apply subgroup reweighting to correct data imbalance, and introduce gate entropy regularization to prevent routing from collapsing onto subgroup attributes, keeping expert utilization both balanced and interpretable. Beyond improving fairness, the routing distribution offers an interpretable view of how subgroups are allocated across experts. Experimental results demonstrate that the proposed approach improves fairness while maintaining competitive predictive performance.
Although recent deepfake detectors achieve high overall accuracy, their errors remain unevenly distributed across demographic subgroups, with real faces from certain groups more often misclassified as fake. Existing fairness-aware detectors typically regularise the entire feature representation, without identifying or controlling the specific components that drive unfair predictions. Such coarse intervention can over-suppress useful forgery cues while leaving demographic structure in component-specific subspaces. To address this, we identify two subgroup-sensitive components: multi-scale spatial features, which encode local facial and forgery patterns, and fine-tuning-induced residual features, which adapt the backbone to the unfair training distribution. We propose FairReL, a fairness-aware representation-learning framework that targets both components with dedicated demographic supervision. FairReL uses an SVD-decomposed foundation-model backbone to isolate the fine-tuning-induced residual representation, and introduces two complementary losses. Group-Conditional Wavelet Decorrelation (GCWD) suppresses subgroup-imbalanced structure across spatial wavelet sub-bands, while Subspace-Localised Mean Alignment (SLMA) aligns subgroup means within each real/fake class in the residual representation. Experiments on FF++, Celeb-DF, DFD and DFDC show that, against the state-of-the-art fairness-aware detector, FairReL improves unseen-dataset AUC by 3.9% while reducing subgroup FPR disparity by 10.2%. Code is available at https://github.com/xiaoman89/FairReL .
Xiaoman Lu, Jiaqi Li, Shuntian Zheng et al.· 0 citations
Deep recommender systems frequently suffer from algorithmic biases that lead to unequal recommendation exposure across demographic groups, particularly disadvantaging underrepresented users. Although various debiasing methods have been proposed, they typically rely on invasive architectural adjustments that disrupt the core algorithm, causing a severe trade-off between fairness and predictive accuracy. To address this, we introduce a novel comparative framework that leverages Generative Adversarial Network (GAN)-based data augmentation to fix the root cause of data imbalance and enhance user-side fairness without sacrificing recommendation performance. Using Neural Collaborative Filtering (NCF) on the MovieLens dataset, we systematically benchmark two distinct augmentation strategies. First, we evaluate decoupled fairness augmentation, which pairs traditional GAN models (CTGAN, GReaT, and TAEGAN) with external preprocessing mitigations. Second, we evaluate intrinsic fairness-aware augmentation, which employs GAN models with embedded fairness mechanisms (TabFairGAN, DECAF, and CFGAN). Performance is evaluated using a comprehensive suite of fairness and recommendation quality metrics. The results demonstrate that GAN-based augmentation significantly improves user-side fairness while preserving overall recommendation performance. Notably, the most effective approach, TabFairGAN, achieves the best trade-off between fairness and accuracy, recording the lowest Uabs of 0.003 and the highest NDCG@10 of 0.811. These findings establish GAN-based data augmentation as an effective, non-invasive strategy for mitigating user-side bias in recommender systems.
Radhofan Azizi Ramdhani, Rita Rismala· International Conference on...· 0 citations
Vision-language models have displayed remarkable capabilities in multi-modal understanding and are increasingly used in critical applications where economic and practical deployment constraints prohibit re-training or fine-tuning. However, these models can also exhibit systematic biases that disproportionately affect protected demographic groups and existing approaches to addressing these biases require extensive model retraining and access to demographic attributes. There is a clear need to develop test-time adaptation (TTA) approaches that improve the fairness characteristics of pretrained models under distributional shift. In this paper, we evaluate how episodic TTA affects fairness in CLIP classification under subpopulation shifts and develop FairTPT, a novel fairness-aware episodic TTA method that jointly minimizes target marginal entropy while maximizing spurious marginal entropy through soft-prompt tuning. We find that standard episodic TTA generally exacerbates disparities between majority and minority groups, that blinding a model to spurious attributes without degrading target performance is inherently challenging, and that excessive blinding can lead to catastrophic forgetting. This model collapse can be prevented by monitoring test-time changes in target loss within the linear regime, while still achieving fairness improvements on reactive data and preserving overall performance. FairTPT outperforms all state-of-the-art episodic test-time debiasing methods and establishes a foundation for robust TTA, which is essential for achieving fairness in practice.
Yoann L. Launay, Parameswaran Kamalaruban, Tom Kempton et al.· 0 citations
Deep learning-based credit scoring systems face three interrelated challenges typically addressed in isolation: unavoidable class imbalance, model opacity, and demographic inequalities. This paper proposes the Explainable Fairness-Aware Deep Learning (EFADL) framework, which is a unified end-to-end pipeline designed to mitigate these risks simultaneously. The EFADL framework integrates three novel components: FC-SMOTE, a fairness-constrained oversampling module that preserves intra-group demographic balance; a multi-objective joint training loss combining focal imbalance correction with differentiable multi-attribute fairness penalties; and a dual-level SHAP module providing both instance-level adverse action explanations and group-level fairness attribution. Extensive experiments on the German Credit, Taiwan Credit, and LendingClub datasets demonstrate that EFADL achieves a better accuracy-fairness trade-off surface. Results indicate a 79% reduction in statistical parity and equal opportunity differences and a 70.5% decrease in maximum intersectional disparity, with a negligible AUC-ROC cost of only 1.2 percentage points. Furthermore, the framework reduces the fairness attribution gap by 71%, which provides evidence that it achieves fairness by suppressing reliance on demographic proxies rather than post-hoc calibration. By delivering stable, economically interpretable explanations, the EFADL framework aligns with the transparency requirements of the EU AI Act and US CFPB guidance and offers a deployable solution for regulatorily-compliant algorithmic lending.
Unknown authors· International Journal of Dat...· 0 citations
Although neural-based machine learning models have received a lot of attention recently, tree-based models such as gradient boosting are competitive for tabular data and therefore remain widely used in various applications of AI. As when using other machine learning predictive models, they can however yield discriminative predictions across demographic groups, due to so-called algorithmic biases. These undesirable phenomena have motivated the emergence of new regulatory frameworks and various AI fairness strategies. While several pre-and post-processing methodologies exist to mitigate such bias on gradient boosting models, only a few in-processing methods have been proposed. To bridge this gap, we introduce OT-FairBoost, a novel in-processing framework that incorporates a Wasserstein-2 distance penalty directly into the objective function of gradient-boosted trees. This OT-based mitigation strategy has been shown to efficiently optimize group fairness criteria such as Demographic Parity and Equalized Odds on neural-based predictions. To adapt this approach for gradient boosting, we extend the sample-wise gradient estimation of the Wasserstein-2 distance between group predictions to discrete distributions and hessian diagonals. We then integrate our approach into the LightGBM training procedure and evaluate it across binary classification, regression, and multi-group sensitive attribute settings. Experimental results in each of these settings demonstrate that OT-FairBoost achieves best accuracy-fairness trade-offs against alternatives.
Veronika Shilova, Abdoulaye Sakho, Younes Boumoussou et al.· 0 citations
As machine learning (ML) algorithms are increasingly used in high-stakes applications, concerns have arisen that they may be biased against certain social groups. Although many approaches have been proposed to make ML models fair, they typically rely on the assumption that data distributions in training and deployment are identical. Unfortunately, this is commonly violated in practice and a model that is fair during training may lead to an unexpected outcome during its deployment. Although the problem of designing robust ML models under dataset shifts has been widely studied, most existing works focus only on the transfer of accuracy. In this paper, we study the transfer of both fairness and accuracy under domain generalization where the data at test time may be sampled from never-before-seen domains. We first develop theoretical bounds on the unfairness and expected loss at deployment, and then derive sufficient conditions under which fairness and accuracy can be perfectly transferred via invariant representation learning. Guided by this, we design a learning algorithm such that fair ML models learned with training data still have high fairness and accuracy when deployment environments change. Experiments on real-world data validate the proposed algorithm. Model implementation is available at https://github.com/pth1993/FATDM.
