Research On Cross-end Joint Personalized Recommendation Based On Time-dynamic Federated Learning

Recommender system enables users to find the most useful items in the most appropriate way.Federated learning enables the user's personalized data to be protected.solve the privacy protection problem of the user's personalized data.Federated recommender system is the interdisciplinary between federated learning and recommender system,which apply the idea of federated learning to recommender system.This thesis makes research on the federated learning and federated recommender system.There are main contents:(1)Aiming at Non-IID,the general problem in federated learning,this thesis proposes a Time Decay Factor based Dynamic Clustered Federated Learning(TDFL).Before local training,every client uses weighted average method,which weights is based on time decay factor,to combine the historical parameters in the local historical parameter list and the global parameters sent by the server;after receiving the parameters uploaded by the client,the server not only aggregates the parameters,but also clusters them to get clustered average parameters,which will send back to their belonging clients,so that the clients can update their local historical parameter list.TDFL solves the Non-IID problem in federated learning.(2)Aiming at the problem that the server can not make personalized recommendation due to the lack of original user data in the traditional federated recommendation system,this thesis proposes a User Embedding Features based Recommender System Model,in order to make the server and client joint personalized recommendation.User Embedding Features are the output results of the Embedding layer,which input is the original user data.The server cannot get the original user data only from User Embedding Features,so the privacy of user data is effectively protected.Therefore,the server can use the User Embedding Features for personalized recommendation,and send the candidate item set to the clients,so that the clients can use the original user data for local personalized recommendation,so as to make the server and client joint personalized recommendation.(3)Because users are generally unwilling to provide user data in the recommender system after they associate with third-party software and log in with a third-party software account,this thesis proposes a Third-Party User Identifier based Recommender System Model.The server cannot get the original user data from the user identifier provided by the third party,which effectively protects the data privacy of clients.Therefore,the server can use the Third-Party User Identifier for personalized recommendation,and send the candidate item set to the clients,so that the clients can use the original user data for local personalized recommendation,so as to make the server and client joint personalized recommendation.Experiments in this thesis show that TDFL is better than the traditional federated learning method.TDFL can be effectively applied to the server and client joint personalized recommendation.