Attention-based federated learning for multi-dimensional personalized recommendation system

Lining Zheng, Hong Chen, Yuanguo Lin, Li Li, Yuanhui Yu, Hefeng Chen, Kangkang Li, Runshan Hu, Shutong Xie

Abstract: Federated learning offers a privacy-preserving solution for personalized recommendation by enabling collaborative model training across multiple clients without sharing sensitive user data. However, existing federated recommendation systems still struggle with three intertwined challenges: capturing multi-dimensional user preferences, modeling complex user–item interactions in distributed settings, and balancing the trade-off between privacy protection and recommendation accuracy. To this end, we propose Attention-based Personalized Federated Recommendation (APFR), a novel framework that integrates attention mechanisms to enhance both personalization and contextual understanding. APFR introduces a personalized score function and user embedding strategy to more precisely represent diverse user preferences, while multi-head attention is employed to capture intricate user–item relationships and contextual dependencies. Furthermore, APFR incorporates differential privacy techniques, such as Laplace and Gaussian noise, to ensure robust privacy protection with minimal accuracy loss. Experimental results on benchmark datasets demonstrate that APFR achieves superior performance in recommendation accuracy and privacy preservation, compared to state-of-the-art methods. The code and datasets are available at https://github.com/lyg2618/APFR.

External IDs:dblp:journals/jksucis/ZhengCLLYCLHX25