Model-Based Debiasing for Groupwise Item Fairness

Yikai Zhang, Yeshaya Adler, Mina Dalirrooyfard, Teresa Datta, Volodymyr Volchenko, John P Dickerson, Yuriy Nevmyvaka

Keywords: Recommender Systems, Item Fairness, Post-processing

Abstract: Recommendation systems are an essential tool for presenting items to users, and hence they are subject to many fairness considerations for users and items alike. Many post processing algorithms exist to handle unfairness in recommender systems; however, they can be very inefficient and not suitable to be used in real time as they need the whole data set to be able to calibrate the recommender system's output. We develop the first model-based group-wise item fairness post-processing algorithm for recommendation systems using a neural network architecture which learns in a data-dependent fashion. Our model adapts and refines itself based on the underlying data without significantly compromising the original utility during the training phase of the recommender system. These performance guarantees are ensured by VC theory and stochastic approximate analysis and we showcase our method's capabilities through experiments on synthetic data.

Submission Number: 3