DualFair: Truthful Illustration Studying at Each Group and Particular person Ranges through Contrastive Self-supervision
Authors: Sungwon Han, Seungeon Lee, Fangzhao Wu, Sundong Kim, Chuhan Wu, Xiting Wang, Xing Xie, Meeyoung Cha
Summary: Algorithmic equity has develop into an essential machine studying downside, particularly for mission-critical Net purposes. This work presents a self-supervised mannequin, known as DualFair, that may debias delicate attributes like gender and race from realized representations. In contrast to present fashions that concentrate on a single kind of equity, our mannequin collectively optimizes for 2 equity standards — group equity and counterfactual equity — and therefore makes fairer predictions at each the group and particular person ranges. Our mannequin makes use of contrastive loss to generate embeddings which might be indistinguishable for every protected group, whereas forcing the embeddings of counterfactual pairs to be comparable. It then makes use of a self-knowledge distillation methodology to take care of the standard of illustration for the downstream duties. Intensive evaluation over a number of datasets confirms the mannequin’s validity and additional reveals the synergy of collectively addressing two equity standards, suggesting the mannequin’s potential worth in truthful clever Net purposes
