Designing fair algorithms has recently appeared as a major issue in machine learning, while it has been studied for long in economics, especially in welfare economics and social choice theory. In this talk, I will discuss two ways in which concepts from this literature can help us design fair recommender systems, and the computational challenges they involve.
The traditional objective of recommender systems is to provide users with recommendations that best satisfy their individual preferences. In this context, existing fairness audits for parity in predictions conflict with the heterogeneity of user preferences. We first study envy-freeness, a preference-based fairness notion from social choice, as a complementary diagnosis of recommender systems, which precludes the unfair situation where some (groups of) users are not given better recommendations when such are available to others.
Recommender systems also impact the producers of recommended items: for example, exposure of content on social media generates revenue for content producers. Therefore, exposure on a platform should be fairly allocated to improve the utility of the worse-off items. In our second contribution, we address the problem of generating fair recommendations by maximizing concave social welfare functions of users’ and items’ utilities. Grounded in cardinal welfare economics, our approach satisfies the properties of Pareto efficiency and maximal redistribution from better-off to worse-off.