Opt-in Transparent Fairness for Recommender Systems

Bjørnar Vassøy, Benjamin Kille, Helge Langseth

Abstract: Recommender systems serve large, diverse userbases and are subject to fairness concerns. Seminal work on consumer-side fairness in recommender systems tends to focus on mitigating the influence users’ demographic information has on their recommendations. Recent work allows the users to specify which information should be concealed, but these methods are often complex and inadvertently deteriorate the recommendation performance of users who reveal all information. We propose a simple, novel approach that interfaces with trained non-fair models and allows users to opt-in for improved fairness without affecting the recommendations given to others. Unlike previous work, the proposed method includes fairness transparently. Quantitative comparison to the state of the art indicates that the proposed method has a favourable fairness/utility tradeoff.

External IDs:doi:10.1007/978-3-031-88708-6_23