Research On Recommendation Algorithm Based On Federated Learnin

In the 5G era,the amount of information on the network is growing exponentially,which brings a huge information burden to users.In order to help users cope with the problem of information overload,recommendation systems have emerged.Recommendation systems use users’ historical interaction data to provide personalized recommendation services,becoming a powerful tool to alleviate the problem of information overload.However,traditional recommendation models require centralized data storage and training,which raises concerns about user privacy.Therefore,it seems more reasonable to implement privacy protection strategies that do not collect user data for recommendations.Federated Learning is a distributed method that allows each client to keep their data locally while collaboratively training a global model.This method has the advantages of protecting privacy data,distributed computing and storage,etc.When combined with the recommendation system,Federated Recommendation Algorithm can always keep user data locally on the client side,protecting user privacy in the model building process.However,the existing Federated Recommendation Architecture is not suitable for practical business applications.In fact,user behavior data is distributed on different application platforms,and advertisers may also place the same products on different platforms.This results in service providers being unable to feed accurate user and product features when training the model.In addition,due to the sparsity and distributed storage of data,the model is difficult to avoid the cold start problem.To solve these problems,this paper mainly studies from two aspects.(1)Federated recommendation model based on three-party cooperation.This paper abstracts and summarizes real-world commercial recommendation scenarios to establish a corresponding model,and proposes a hypothesis that accurate recommendations require perfect cooperation between service providers,users,and advertisers.Therefore,this paper proposes a federated learning algorithm based on the three-party cooperation of User-Server-Advertiser(FedUSA).For user data,this algorithm only stores user features on user devices and aggregates and trains user embeddings on user devices,and then sends user embeddings to the interacting service provider.Thus,the service provider can obtain a more accurate user embedding representation without collecting user raw data.For item data,the advertiser platform aggregates and trains item embeddings and sends them to the collaborating service provider.Thus,without exchanging commercial data,the service provider can obtain a more accurate item embedding representation among service providers limited by business competition and privacy protection clauses.When the service provider receives both user embeddings and item embeddings,the model can be updated and joint training can be completed.Based on federated cooperation,FedUSA learns user and item features while ensuring data privacy,making the model more capable of learning richer interaction information.To verify the effectiveness of the method,this paper conducts multiple comparative experiments on multiple real datasets,and the experimental results show that this method improves the recommendation effect.(2)Heterogeneous federated recommendation model based on meta-learning.In the context of heterogeneous federated recommendation models,where three parties participate in training,the data is independent and identically distributed but does not follow the same sampling method(Non-IID).The problem of data heterogeneity can lead to a significant decrease in model accuracy.To adapt the FedUSA model to heterogeneous scenarios,this paper proposes a Federated Meta Embedding(FedME)based on meta-learning under the FedUSA federated recommendation model framework.First,the heterogeneous data is partitioned into Hot ID or Cold ID based on whether the number of instances exceeds or is below a certain threshold,and the recommendation task is completed in two phases.In the federated phase,the goal is to learn richer semantic information for Hot ID through FedUSA collaboration.In the meta-learning phase,each Cold ID is regarded as a learning task,and the parameters learned in the federated phase initialize the meta-learner,which continues to train the embedding for Cold ID.Experimental results demonstrate that FedME can improve the predictive performance of recommendation in a heterogeneous environment,particularly in enhancing the accuracy of predicting Cold ID.