A Personalized Federated Matrix Decomposition Recommendation Algorithm Based on Meta-distillation

The heterogeneity of client-side data in federated recommender systems poses challenges in training a unified model that can cater to the individual needs of each user for recommendation performance. To address this issue, we propose a recommendation algorithm for personalized federated matrix decomposition called HE-MD-PFedMF (Personalized Federated Matrix Factorization Recommendation Algorithm Based on Homomorphic Encryption and Meta-Distillation). HE-MD-PFedMF incorporates user bias item parameters, project bias term parameters, and score record average parameters into the model training process, thereby accommodating variations in project scoring standards among different users. By integrating meta-learning and knowledge distillation techniques into the federated matrix factorization model, our algorithm aims to enhance personalized recommendation accuracy by leveraging user terminal preference data comprehensively. Additionally, we employ homomorphic encryption technology to encrypt gradient parameters, mitigating concerns regarding privacy leakage during the transmission of model parameters. Experimental analysis demonstrates that HE-MD-PFedMF yields lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values compared to prominent baseline algorithms, thereby achieving superior recommendation accuracy.