Balancing privacy, fairness and model utility in deep learning recommender systems

Abstract

Recommender systems play a central role in connecting users to content, shaping access and personalization in digital platforms. Despite their utility, these systems face persistent challenges related to fairness and privacy. This thesis, Balancing Privacy, Fairness, and Model Utility in Deep Learning Recommender Systems, presents five complementary works that address these challenges from multiple perspectives. The first two works focus on the long-tail recommendation problem, where items with limited interactions are systematically under-recommended. To address this, I propose biased user history synthesis, a sampling and augmentation technique that improves user representations, enhancing personalization while boosting performance across both head and tail items. This approach is further extended through a tailored sampling strategy within a mixture-of-experts framework, yielding additional performance gains. The third work tackles machine unlearning in recommender systems, motivated by emerging privacy regulations. I introduce an Enhanced Exact Machine Unlearning (EEMU) method based on meta-learning, enabling efficient and precise removal of user data while preserving recommendation quality. The fourth and fifth works explore linguistic diversity and fairness in Arabic-speaking contexts. By incorporating dialectal signals from user reviews, I demonstrate improvements in recommendation performance and reveal systematic disparities across dialect groups, highlighting fairness challenges in multilingual settings. Together, these contributions advance the development of recommender systems that are more effective, fair, and privacy-compliant.