A Re-ranking Approach for Two-sided Fairness on Recommendation Systems

Abstract

The filter bubble problem has long constrained users of recommender systems from using it freely. Two stakeholders of the recommendation system, which refer to the content consumer, and the content provider, are disturbed by the meaningless repeating of few high-frequency contents. While most previous work concerns the fairness issue of recommenders from one side, in this paper we provide a new lightweight approach through a re-ranking method increasing fairness for both sides. Experiments on 2 datasets and 4 existing models demonstrate that our proposed algorithm can reduce unfairness and increase overall accuracy. The time complexity for our approach is linear to the total user amount for each user. And it fits all existing recommendation systems that generate a rank score.