Achieving Fairness via Post-Processing in Web-Scale Recommender Systems

Building fair recommender systems is a challenging and extremely important area of study due to its immense impact on society. We focus on two commonly accepted notions of fairness for machine learning models powering such recommender systems, namely equality of opportunity and equalized odds. These measures of fairness make sure that equally "qualified" (or "unqualified") candidates are treated equally regardless of their protected attribute status (such as gender or race). In this paper, we propose scalable methods for achieving equality of opportunity and equalized odds in rankings in the presence of position bias, which commonly plagues data generated from recommendation systems. Our algorithms are model agnostic in the sense that they depend only on the final scores provided by a model, making them easily applicable to virtually all web-scale recommender systems. We conduct extensive simulations as well as real-world experiments to show the efficacy of our approach.