Research And Implementation Of Video Recommendation System Based On Edge Computing And Federated Learnin

With the booming development of various video applications on the Internet,video recommendation systems have been rapidly revolutionised.Video recommendation systems can analyse users’ historical viewing behaviour,extract their preferences and provide them with personalised video recommendations.However,most of the mainstream video recommendation systems currently use cloud computing architecture,resulting in excessive computational load on cloud servers and high network latency,which affects the user experience.At the same time,with the enactment of privacy protection laws and increased awareness of user privacy and security,recommendation systems may no longer be allowed to collect private user data to train recommendation models in the future.Therefore,how to provide personalised recommendation services while protecting the privacy of user data becomes an increasingly important issue.To address this problem,edge computing and federated learning technologies offer new ideas.Edge computing leverages the computing and storage capabilities closer to the user at the edge,keeping data at the edge to protect user data privacy and provide faster responses.Federated learning,on the other hand,can use the data at the edge to train models without uploading private user data,enabling personalised recommendation services with security in mind.Therefore,this paper proposes a video recommendation system based on edge computing and federated learning,with the following main work:(1)To address the problem of traditional cloud-centric recommendation systems collecting user privacy data,this paper proposes a video recommendation model based on federation learning called Fed-DIN(Federated Deep Interest Network).The model introduces a federation learning framework based on the deep interest network recommendation model,and addresses the global model performance degradation caused by the non-independent homogeneous distribution of data at each edge of the traditional federation learning,adding an attention mechanism to the federation averaging algorithm to reduce the impact of local model personalization on the global model performance,and introducing a cyclic learning rate to accelerate model convergence.The experimental results show that on the public movie dataset Movie Lens-1M,the performance of the Fed-DIN model in this paper improves by2%-4% over the traditional federal recommendation model,with the dual advantages of improving the recommendation effect and privacy protection.(2)To address the problems of high network latency and high cloud server load in the traditional cloud-centric recommendation system architecture,this paper designs and implements a video recommendation system architecture based on edge computing and federation learning.The architecture divides the recommendation process into two stages: recall and sorting,and designs various recall strategies according to different needs.Finally this paper details the implementation process of the core modules through the steps of requirement analysis,overall architecture and functional module design,and verifies the effectiveness of the system in this paper through testing.