A General Framework for Fairness in Multistakeholder Recommendations
This addresses fairness issues for buyers, sellers, and platforms in multi-sided recommender systems, representing an incremental improvement over existing methods.
The paper tackled the problem of balancing multiple stakeholder objectives in recommender systems, where traditional offline methods with global seller coverage constraints can lead to low utility for some buyers; they proposed a real-time personalized framework that incorporates seller coverage and individual buyer objectives, achieving provable theoretical bounds and empirical validation on real-estate marketplace data.
Contemporary recommender systems act as intermediaries on multi-sided platforms serving high utility recommendations from sellers to buyers. Such systems attempt to balance the objectives of multiple stakeholders including sellers, buyers, and the platform itself. The difficulty in providing recommendations that maximize the utility for a buyer, while simultaneously representing all the sellers on the platform has lead to many interesting research problems.Traditionally, they have been formulated as integer linear programs which compute recommendations for all the buyers together in an \emph{offline} fashion, by incorporating coverage constraints so that the individual sellers are proportionally represented across all the recommended items. Such approaches can lead to unforeseen biases wherein certain buyers consistently receive low utility recommendations in order to meet the global seller coverage constraints. To remedy this situation, we propose a general formulation that incorporates seller coverage objectives alongside individual buyer objectives in a real-time personalized recommender system. In addition, we leverage highly scalable submodular optimization algorithms to provide recommendations to each buyer with provable theoretical quality bounds. Furthermore, we empirically evaluate the efficacy of our approach using data from an online real-estate marketplace.