A Dual-Constraint Optimization Approach for Algorithmic Fairness
| Type of publication: | Conference paper |
| Citation: | Fatoretto_IEEECEC_2026 |
| Publication status: | Published |
| Booktitle: | 2026 IEEE Congress on Evolutionary Computation (CEC) |
| Year: | 2026 |
| Month: | June |
| Pages: | 1-6 |
| Publisher: | IEEE |
| DOI: | 10.1109/cec69861.2026.11727425 |
| Abstract: | A robustly fair model must mitigate systematic biases without sacrificing predictive power. While various metrics exist to quantify disparate outcomes, a central challenge remains: navigating the inherent friction between conflicting fairness definitions, such as Independence and Separation. This study investigates a mathematical programming approach to regulate the training of logistic regression models by enforcing multiple, simultaneous fairness constraints. We utilize decision boundary covariance to implement limits on demographic disparities and error-rate imbalances. Our methodology employs Pareto-based analysis to explore the frontier between accuracy and multiple fairness objectives across a range of thresholds. We validate this framework using the Adult Income dataset alongside Arrhythmia and Boston benchmarks. Our results, supported by Hyperarea (HA) and Spacing (SP) indicators, demonstrate that joint constraints provide a more stable calibration of fairness than single-objective approaches. By making these trade-offs explicit, we provide a mechanism for clinicians to select models aligned with specific ethical and institutional priorities. |
| Main Research Program: | AI for Everyone |
| Keywords: | Adult Income dataset, algorithmic fairness, demographic disparities, fairness-accuracy trade-off, logistic regression, Pareto front |
| Projects: |
FAIRMI |
| Authors: | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
Notes
|
|
|
|
|