Federated learning for recommender systems
Συνεργατικές τεχνικές μάθησης σε συστήματα συστάσεων
View/
Keywords
Federated learning ; Recommender systems ; Personalized recommendations ; Privacy-preserving machine learningAbstract
Recommendation systems are central to many digital platforms, enabling personalized content delivery in domains such as e-commerce, streaming services, and social media. However, conventional recommendation models often require centralized access to user data, which raises significant privacy and security concerns, especially under modern data protection regulations such as the GDPR. Federated Learning (FL), a privacy-preserving collaborative training paradigm, addresses these issues by keeping user data localized on client devices while aggregating only model updates on a central server. This makes it highly suitable for personalized recommendation tasks, forming the basis of Federated Recommendation Systems (FedRS). This thesis explores the application of advanced deep learning models in a federated setting for the task of personalized item recommendation. Multiple federated algorithms are evaluated, including the standard Federated Averaging (FedAvg), the Weighted Federated Averaging (WFedRec), and the proposed clustering-enhanced method, CWFedRec. The proposed CWFedRec introduces client clustering based on semantic user preferences and incorporates personalized bias vectors generated using sentence transformers. These algorithms aim to address challenges specific to FedRS, including data heterogeneity, limited communication bandwidth, and the curse of dimensionality in high-dimensional embedding spaces.