GenFair: A Genetic Fairness-Enhancing Data Generation Framework

Bias in the training data can be inherited by Machine Learning models and then reproduced in socially-sensitive decision-making tasks leading to potentially discriminatory decisions. The state-of-the-art of pre-processing methods to mitigate unfairness in datasets mainly considers a single binary sensitive attribute. We devise GenFair, a fairness-enhancing data pre-processing method that is able to deal with two or more sensitive attributes, possibly multi-valued, at once. The core of the approach is a genetic algorithm for instance generation, which accounts for the plausibility of the synthetic instances w.r.t. the distribution of the original dataset. Results show that GenFair is on par or even better than state-of-the-art approaches.