Fair age prediction through ILP-based multi-attribute training subset selection

As algorithms play an increasingly prominent role in decision-making, concerns have grown over their disproportionate impact on minority groups. This has led to a surge of interest in algorithmic fairness, a field dedicated to embedding fairness principles into algorithm development. Motivated by these challenges, we extend a fair coverage problem, termed Multi-Attribute Fairer Cover, which aims to identify a maximally fair subset for training purposes. In this study, we conduct new experiments using an Integer Linear Programming ( ilp ) model to create fair training instances for an age regression task. We compare these fair instances with randomly selected samples from the UTKFace dataset, evaluating the performance with classical convolutional neural network, state-of-the-art convolutional model, and attention-based regressors. Results show a significant error reduction in the regression model when using fair selection compared to random selection, both in general scenarios and when filtered by gender and race, including minority groups. • We adapt the fair subset selection for use with the tabular datasets. • We introduce an ilp model to select fair UTKFace subsets. • We present experiments where our fair selection outperforms random selection. • Our approach improves performance across various regression models.