Research On Intelligent Recommendation Technology Based On Federated Learning

With the continuous development of the Internet,the number of types of information available to users is increasing and the phenomenon of ’Information Overload’ has emerged.It is in this context that the recommendation system is widely developed and has become one of the popular artificial intelligence technologies in industry and academia.At the same time,the global community is paying more attention to data privacy,and a large amount of private data is not allowed to be spread freely,leading to the phenomenon of ’Data Silos’ in key industries such as finance and healthcare,and a large amount of valuable data is difficult to be used for research in a timely manner.Federated learning provides an idea to solve the above contradiction,which provides reliable machine learning training while protecting data security.In this thesis,we focus on the intelligent recommendation technology based on federated learning,and propose a recommendation method based on federated learning with neural collaborative filtering structure and an improved differential privacy federated learning recommendation method based on it.In the neural collaborative filtering structure recommendation method based on federated learning,the Neu MF method is used to predict the recommendation effect,and the federated learning algorithm Fed Avg is used for gradient averaging and weighted aggregation to realize the training mode of training data only locally and gradient uploading to the central server.In the improved modified differential privacy based federal learning recommendation method,differential privacy is used to add Gaussian mechanism based noise to the interacting gradients,which effectively prevents the back-inference of the dataset due to gradient leakage.The Fed Prox improved algorithm of Fed Avg combined with NCF is used to optimize the recommendation effect of the federal learning neural system filtering model.The experimental results show that the neural collaborative filtering algorithm based on federated learning constructed in this paper can achieve the effect close to the neural collaborative filtering recommendation under the premise of protecting privacy security,and the difference of HR@10 under the Neu MF framework is only 2.37%.After the introduction of differential privacy,the recommendation performance drops by only 1%,which verifies the feasibility of differential privacy.After adopting the improved algorithm Fed Prox,the recommended evaluation index is improved by more than 2% in most cases,indicating that the local heterogeneous problem has a convergence effect,and the improvement of the algorithm is effective.