Improving Fairness of Recommender Systems

Abstract

Quals talk:

The deployment of recommender systems is becoming ubiquitous to help customers navigate the overwhelming information available on platforms such as social media, online shopping, and streaming services. However, biased recommender systems may put people from racial and ethnic minority groups at higher economic and social risk, particularly in areas like insurance and loan. Therefore, fairness is a crucial consideration in the development of these systems and is under increasing scrutiny. In this talk, we propose an algorithm that enhances the fairness of collaborative recommender systems, which are known to be biased by factors such as the tyranny of the majority, by introducing content-based factors into the process. Our algorithm distinguishes itself from traditional hybrid recommender systems by employing a variant of Latent Dirichlet Allocation (LDA) to recharacterize products in a way that the learned “genres” caters to the personalized needs of customers, ensuring better performance for minority groups.