Statistical Learning in Recommendation Systems: Personalized Ranking and Decentralized Learning Systems

Recommendation systems are integral to various domains, including content-based platforms and e-commerce. Despite extensive research efforts in designing diverse recommendation systems in recent years, popular approaches based on explicit ratings often struggle to perform effectively in practical recommendation-making scenarios. This dissertation comprises three main sections. Firstly, we will present an overview of recommendation systems from a statistical perspective. Secondly, we propose a personalized ranking-based model that utilizes pairwise preference information from explicit feedback, demonstrating superior performance compared to conventional approaches in real-world settings. Finally, addressing growing concerns over privacy and data safety within recommendation systems, we introduce a novel decentralized federated learning framework. This framework operates without relying on centralized data aggregation or costly model merging procedures.