Research On Recommendation Algorithm Based On Network Representation Learning

With the rise of computer and mobile communication technology,accessing information has become more convenient.However,this has also led to an overwhelming amount of data generated,resulting in information overload.Recommendation algorithms have been developed to help filter and screen information,providing users with content tailored to their interests.These algorithms have been successfully applied in various fields such as e-commerce,video and audio entertainment,and social networks.In recent years,network representation learning technology has gained popularity due to its exceptional node representation ability,providing a new direction for the research of recommendation algorithms.However,the nodes in the complex network system in the real world usually contain a variety of different types of relations that interact with each other,and simple single-layer network is not enough to describe them accurately.In addition,nodes usually contain additional attribute information,and how to make good use of these auxiliary information to improve the performance of the algorithm has become a major challenge in current research.To solve the above problems,based on multi-layer network and attribute network,this dissertation uses network presentation learning technology to enhance the performance and scalability of the recommendation algorithm.The main contributions are as follows:(1)This dissertation proposes a multi-layer network representation learning model called MNESE,which integrates structural entropy information to enhance the performance and scalability of the recommendation algorithm.The proposed model considers the local and global structural information of nodes comprehensively,using node local structural entropy and network standard structural entropy to guide the intra-layer and inter-layer migration process.The skip-gram model is then used to learn the node vector representation based on the sampled node sequence.Finally,the similarity of nodes is calculated using the embedding vector of nodes,and the recommendation task is completed based on node similarity.Link prediction experiments were conducted on five real data sets,compared with the mainstream model,the performance of the proposed method is improved by 5% to 10%.This proves the validity of the random walk guidance information.(2)In addition,a recommendation algorithm combining node attribute information is proposed for attribute networks.The algorithm uses an automatic encoder to upgrade or reduce the dimension of the node attribute vector,making it consistent with the embedding dimension of the node.The node local structure entropy information is then used to guide the random walk process to enrich the semantic information of the node sequence.The skip-gram model is improved by fusing the node attribute vector to join the training process of node embedding.Finally,the node vector representation is applied to the recommendation task.Link prediction experiments were carried out on three real attribute network data sets,and the proposed algorithms were found to improve node representation ability and achieve excellent results in the recommendation task.