Research On Recommendation Method Based On Model Fusion In Cloud-edge-end Scenarios

In recent years,the vigorous development of e-commerce,short video and other applications has promoted the rapid innovation of recommender systems.Recommender system mines user preferences through user’s historical interaction logs and portrait information,and recommends personalized items for users.However,the current mainstream recommender systems generally follow a cloud-centered architecture.The service requests are sent to the cloud server,which completes the entire modeling and calculation process and returns the recommended results to users.As the scale of user’s logs increases and the data modalities become more diverse,the computing load on cloud servers increases,and network latency increases,which affects the user experience.In addition,as far as recommendation methods are concerned,most of the current recommendation methods mainly focus on mining users’ explicit preferences,while ignoring users’ implicit preference information,and lack of exploration of users’ micro-behavior information,resulting in poor recommendation quality.In view of the above issues,this thesis uses the model fusion method to study the recommender system in the cloud-edge-end scenarios.The main work is as follows:Aiming at the problems of high network delay and heavy cloud server load in the traditional cloud-centered recommender system architecture,this thesis proposes a recommender system architecture for cloud-edge-end scenarios.The system architecture is mainly divided into three layers.First,the group recommender system deployed on the cloud server realizes group recommendation by integrating regional user feature preferences,and caches the recommended results in the edge server.Then,the personalized recommender system deployed on the edge server recommends the filtered cached items to the user by mining the micro-behavioral characteristics of the target user,and finally the user terminal presents the recommended result.Benefit from a two-stage recommender model,most of the user’s requests will be satisfied at the edge,solving the problem of heavy load and high latency in the traditional mode.Aiming at the problem that traditional cloud-centered recommendation methods are difficult to integrate regional user feature preferences,this thesis proposes a cloud server-oriented group recommendation method based on feature preference fusion.Considering the particularity of cloudedge-end scenarios,users covered by edge servers are regarded as special groups.By integrating knowledge graph embedding technology,attention mechanism and recurrent neural network,the individual preferences and group fusion preferences of regional users are mined to realize group recommendation,and the recommended results are cached in edge nodes.This method solves the problem that the traditional cloud-centered recommendation method is difficult to integrate regional group preferences and realize group recommendation.Aiming at the problem that traditional personalized recommendation methods focus on users’ short-term preferences and ignore users’ long-term preferences and micro-behavior characteristics,this thesis proposes a personalized recommendation method based on micro-behavior mining for edge servers.This method first introduces a self-attention network to mine user preference dependencies,then integrates a recurrent neural network to model the evolutionary representation of users’ dynamic preferences,and fuses users’ long-term and short-term preferences based on similarity theory.In addition,this method fully considers and quantifies the preference strength contained in different micro-behaviors of users,and integrates behavior preferences to realize recommendation.The experimental results verify the effectiveness of the proposed method.Based on the above theories and methods,this thesis constructs a recommender prototype system based on model fusion in the cloud-edge-end scenarios.The prototype system is introduced in detail through the steps of requirement analysis,outline design,detailed design,and concrete realization,which realizes the combination of theory and application.In addition,simulation experiments are designed to verify the effectiveness of the architecture proposed in this thesis.