Aplicação do aprendizado por transferência para otimização de sistemas de recomendação federados.

Abstract

This study investigates the optimization of Federated Recommendation Systems (FRSs) by applying Transfer Learning techniques. Facing challenges posed by high communication costs and the necessity to preserve user data privacy, we propose an innovative methodology that leverages pre-training and knowledge transfer strategies. This research uses well-known datasets such as MovieLens and Netflix Prize to explore knowledge transfer from central ized recommendation systems to a federated approach. Techniques such as PCA (Principal Component Analysis) and Word2Vec are examined in conjunction with federated training to evaluate and utilize the generated embeddings in the federated training process. The findings demonstrate significant reductions in communication costs and the number of users required for effective model training without compromising recommendation accuracy. Additionally, a case study focusing on compliance with the General Data Protection Regulation (GDPR) highlights the practical relevance of the developed techniques for data protection-compliant recommendation systems. This work contributes to the field of Federated Recommendation Systems by showing how integrating Transfer Learning with federated learning can optimize FRSperformance, with practical implications for developing privacy-preserving recommen dation systems in the current technological landscape.