Research On Personalized Service Recommendation Method Based On Federated Learning

With the rapid development of Internet technology in recent years,more and more enterprises have used personalized recommendation technology in their Internet products.Federated learning technology is an effective solution to the problem of privacy protection in personalized recommendation systems.This paper focuses on solving the problem of practical application of the longitudinal federated factor decomposer algorithm in the personalized recommendation field of federated learning due to its low training efficiency and poor performance in the face of massive sparse data.This paper first describes a practical scenario in personalized service recommendation.In order to improve the low training efficiency of the non-sampling traditional longitudinal federation factor decomposer in this scenario,a high-efficiency non-sampling factor decomposer service recommendation algorithm Fed-ENSFM based on the longitudinal federation learning framework is proposed.On the other hand,this paper considers a specific CTR prediction scenario.In order to solve the problem of poor performance of the traditional vertical federation factor decomposer in the face of massive sparse data under this scenario,a personalized service recommendation algorithm Fed-FFM based on the vertical federation feature domain awareness factor decomposer is proposed based on this scenario.Finally,according to the relevant research content,a federated learning movie recommendation system is designed and implemented.The specific research contents of this paper are as follows:(1)In this paper,the efficient non-sampling factor decomposer(ENSFM)is combined with longitudinal federated learning,and the non-sampling federated recommendation algorithm Fed-ENSFM is proposed.Finally,the experimental results show that the implementation efficiency of this algorithm is significantly faster than Fed FM.(2)This paper combines the field perception factor decomposition machine(FFM)with the longitudinal federation learning,and proposes the longitudinal federation CTR prediction algorithm Fed-FFM,which is more suitable for the sparse sample scenario.Finally,the accuracy of the algorithm in the sparse longitudinal federation recommendation scenario is verified by theoretical analysis and experiments.(3)Based on the technology of vertical federation learning and web front-end,this paper designs a movie recommendation system under the vertical federation scenario.The system adopts the B/S architecture,encapsulates the encryption gradient uploaded by participants in the vertical federation recommendation,the encryption and decryption of homomorphic encryption,and forms a specific functional module,and builds a prototype system based on these functional modules.The experiment shows that the training efficiency of Fed-ENSFM is significantly better than that of Fed FM,the traditional federal factor decomposer,on the premise of ensuring the performance of the model.Compared with centralized FFM,Fed-FFM protects the user’s privacy and improves the model accuracy.This paper also implements a vertical federated learning movie recommendation system based on B/S architecture to intuitively display the application process of the vertical federated learning recommendation system in real business scenarios.