Research On Recommendation Method Based On Deep Learning In Cloud-Edge-End Scenarios

With the rapid development of Internet,recommendation systems have been widely used as a key technology in popular applications such as e-commerce and short videos.Recommendation systems aim to discover users’ preference features through data such as users’ historical behavior logs and portraits,and recommend items or services that meet users’ needs.However,with the rapid iteration of communication technology,the expanding data scale leads to the increasing computational cost and network latency of traditional recommendation systems based on cloud computing,which directly affects the user experience.In addition,as far as recommendation methods are concerned,more and more research attempts to alleviate the increasingly severe data sparsity problem through multi-feature approaches,including historical behavior,social networks and personal preferences.However,these methods often treat high-order nonlinear relationships between users as linear relationships,which may result in a lack of semantic features.Besides,these methods mostly ignore the degree of user preference for items at different moments when mining preference features from behavior sequences,which may lead to poor recommendation quality.In view of the above analysis,this thesis focuses on studying the recommendation based on deep learning in the cloud-edge-end scenarios.The main work is as follows:Firstly,this thesis proposes a service matching method based on group preferences and service representation learning for edge caching.By combining edge computing,services are deployed on the edge side close to users to reduce network latency and computational overhead.The method first uses deep embedding clustering technology to map user embedding features to low-dimensional hidden spaces and realizes group division based on this space.Then,gated recurrent units are used to mine group preference features.In addition,a graph attention network is used to learn high-dimensional sparse service representation vectors,which can obtain effective service representation features.Finally,candidate service matching is completed by fusing group preference features and service representation features.Experimental results verify the effectiveness of the proposed method in improving cache hit rates.Secondly,this thesis proposes a recommendation method based on multi-feature fusion by analyzing users’ interaction behaviors and user-item interaction behaviors.This method first uses a hypergraph to model nonlinear interaction relationships between users and mines users’ neighborhood features through a graph convolutional network.Then,long short-term memory networks are used to mine long-term dependencies in users’ historical item interaction sequences,and an attention mechanism is introduced to fuse hidden layer outputs of long short-term memory networks at different moments.Finally,personalized recommendations are implemented based on users’ neighborhood features and users’ preference features.Experimental results verify the effectiveness of the proposed method.Lastly,based on the above theories and methods,this thesis constructs a recommendation prototype system based on deep learning in cloud-edge-end scenarios.Through requirement analysis,outline design,detailed design,and prototype system demonstration,the implementation process and operation flow of the prototype system are described in detail,which verifies the effectiveness of the proposed methods.