Research On Recommendation System Technology Based On Vertical Federated Learning

The development of recommendation system is inseparable from large-scale and high-quality data.Many related companies share data directly to improve the performance of their recommen-dation systems.However,with the gradual improvement of user data privacy awareness and data protection laws,the data privacy security issues of multi-party recommendation systems have at-tracted more and more attention.Focusing on the issues of data privacy protection in multi-party recommendation systems,this paper proposed an efficient recommendation system based on verti-cal federated learning by using federated learning technology,which improves the data privacy and reduces the cost of using privacy protection technology in multi-party recommendation systems at the same time.This paper starts from the data sharing method of the multi-party recommendation systems,and analyzes the data privacy security issues of the participants,and then introduces the verti-cal federated learning technology to improve the data privacy security of the multi-party recom-mendation systems.Then,this paper studies the efficiency of the above methods and proposes the Base-Asynchronous Vertical Federated Learning(Base-AVFL)method.After analyzing the reasons for the decline of the model performance caused by the Base-AVFL method,this paper further proposes Predictor Compensation Asynchronous Vertical Federated Learning(PC-AVFL)which shortens the training time under the premise of ensuring data security while bringing no loss on recommendation accuracy.Aiming to solve the issue of low communication efficiency in the system,this paper uses the approximate sparseness of intermediate data to propose Double End Sparse Compression(DESC)method to compress the communication data between participants.Combining the above two methods,this paper finally proposes a recommendation system based on vertical federated learning under the PC-AVFL-DESC method.Based on the centralized recommendation system,this paper redesigns the recommendation system based on vertical federated learning,builds a test platform environment and integrates the above system into software.Finally,this paper conducts comparative experiments on two public recommendation data sets and three basic recommendation models.The experimental results show that the recommendation system based on vertical federated learning under the PC-AVFL-DESC method guarantees the data security of the participants and compared with the centralized method,the biggest impact is only 1.98%under the evaluation of MSE,AUC and F1.Compared with the method based on Synchronous Vertical Federation Learning(SVFL),the PC-AVFL-DESC method proposed in this paper reduces the average system convergence time by 35.11%and the system communication volume by 76.14%.The above experimental results fully prove the methods pro-posed by this paper is practical and effective.