Fairness in Group Recommender Systems Using Variational Autoencoders

Abstract

Recommender systems are integral to enhancing user experiences on platforms like Amazon and Netflix by providing personalized suggestions. However, these systems often face significant fairness challenges, particularly in group settings where diverse preferences must be aggregated. This thesis explores the use of Variational Autoencoders (VAEs) to improve fairness in group recommendations. By introducing stochastic elements into the VAE framework, the research aims to generate diverse and equitable recommendations. Extensive evaluations using the MovieLens 20M dataset demonstrates that incorporating noise during the recommendation process significantly enhances fairness with a minimal impact on ranking quality. The study identifies the Hybrid aggregation method paired with uniform noise as the optimal tradeoff, balancing group satisfaction, dissatisfaction, and ranking quality. This work contributes to developing more inclusive group recommender systems, offering practical guidelines for implementation and paving the way for future advancements in fair and effective recommendation algorithms.