Reducing Recommendation Inequality via Two-Sided Matching : A Field Experiment of Online Dating

Leading recommender systems may recommend only a small fraction of users on the dating platform since the algorithms often exploit popularity and similarity that reinforce preference homogeneity and assortative mating in the marriage market. We apply a stylized matching model in economics to the existing algorithms to reduce inequality, and we evaluate the proposed method by a large-scale field experiment through a dating app. Our recommender improves predictive accuracy and reduces inequality, leading to substantially more matched couples than the other two competing algorithms. In particular, male users assigned to our novel recommender are four times more likely to receive females' responses. We improve several inequality measures: There are far fewer superstars promoted by our algorithm, and the distribution of the recommendation count is more even than the other two algorithms. Our algorithm also yields the highest coverage rate