Year: 2026
Type: article
Abstract: Ensuring fairness in machine learning models is a fundamental challenge, particularly because of the redlining effect, in which proxy variables highly correlated with sensitive attributes unintentionally reproduce bias. Neural Network fairness-aware strategies often fail to explicitly penalize such correlations, applying dropout on hidden layers, which proves insufficient for controlling proxy discrimination. This study introduces the Statistical Dependency Regularization (SDR), a novel in-processing regularization approach for Multi-Layer Perceptrons (MLPs) that penalizes input-layer weights proportionally to the correlation between features and sensitive attributes. Dependence is primarily quantified through Chatterjee’s correlation coefficient, which is robust to noise and capable of capturing complex, nonlinear, and non-monotonic relationships. Comprehensive experiments, including ablation studies, across four widely adopted fairness benchmarks, optimizing six combined objectives involving multiple predictive performance with widely adopted fairness metrics, under hyperparameter search and resampling evaluations, compared through a suitable significance test, indicate that SDR effecively mitigates redlning effect, producing competitive results on multiple scenarios. Hence, by directly penalizing proxy effects, SDR provides a principled and effective contribution to fairness-aware neural networks, showing substantial potential for real-world deployment in sensitive decision-making systems. We also provide a theoretical analysis to support the intuition behind the proposed approach, showing how the SDR progressively reduces the model’s reliance on features correlated with the sensitive attribute and, in the limit, the predictor satisfies statistical parity under joint-independence assumption.
Source: Expert Systems with Applications
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: hybrid
Estimated APC paid: $3,500
OpenAlex ID: W7203752928
