Achieving Intersectional Algorithmic Fairness by Constructing a Maximal Correlation Latent Space

Recent developments in algorithmic fairness started to investigate the interaction between multiple sensitive information through an intersectional perspective. We introduce a new definition of intersectional fairness based on a multivariate extension of the Generalized Disparate Impact (GeDI). Our approach leverages a neural network to transform multiple protected groups into a univariate latent space that maximizes correlation with the target, effectively capturing unfairness across all potential subgroups even with limited data samples. Empirical evaluations on several benchmarks demonstrate that our method can be effectively used as a loss regularizer during neural network training, offering stronger performance guarantees compared to existing intersectional statistical parity definitions while also allowing to manage continuous inputs and targets.