Federated Recommendation System Based On Personalized Federated Learning

The emergence of recommendation systems has solved the problem that in the Internet era,it is more and more difficult for users to make choices due to the complicated information.However,in the training process of recommendation systems,a new contradiction has emerged: the amount of data from a single data owner is often insufficient to support the training of recommendation system models,and multiple data owners cannot share data directly due to the limitation of user privacy protection.The emergence of federated recommender systems solves this problem,but also creates new problems: 1.when a new participant tries to build a recommender system,the lack of historical data support makes the recommender system’s recommendations ineffective,i.e.,the "cold start" problem;2.the data owned by the parties involved in the training of federated recommender systems often do not obey the independent homogeneous distribution,which leads to the recommendation effect of the trained recommendation system not reaching the expectation.In this paper,we analyze the causes of these problems and propose a federal recommendation system based on federal personalization.This paper first proposes a federal personalization method to prevent catastrophic forgetting.The method introduces the plastic weight consolidation method in lifelong learning,and regularizes the model parameters by plastic weight consolidation when the federal learning client performs local model training,and uses sliding average to balance the knowledge learned by the client model from the public model and the client in the previous round.Knowledge transfer is also performed between the public model and the client model,and between the client model of the previous round and the client model of the current round,which prevents catastrophic forgetting and ensures the prediction accuracy of the public model while improving the federal personalization effect.Experiments show that the proposed method improves the average prediction accuracy of client models by 4.3%,22.2%,5.7% and 25.8% in the four scenarios,respectively.Then,based on this federal personalization approach,a recommendation system based on personalized federal learning is constructed in this paper.The system is divided into a data layer,a feature layer,a federation layer and an application layer,which are responsible for data processing,feature fusion,federation learning organization and practical application of the recommendation system,respectively.Finally,based on the above recommendation system,this paper proposes an improvement method to save system communication resources during the training process.The comparison experiments on different recommendation system models and data sets show that: 1.The recommendation system designed based on the personalized federated learning method proposed in this paper improves the recommendation accuracy by 4.2%;2.the recommendation system constructed in this paper solves the "cold start" problem of the recommendation system for new participants to a certain extent;3."The improved method of saving communication and other resources not only saves some system communication resources,but also improves the recommendation accuracy of the public model of the recommendation system to a certain extent.