Research On Social Recommendation Method Based On High-order Node Information Fusion

With the rapid development of e-commerce and social media platforms,recommendation system has become an irreplaceable tool in many fields.Users use recommendation systems to ease the problem of excessive information and explore what they are interested in from the vast ocean of items;With the help of the recommendation system,the merchants can accurately find the target users for their products.To achieve this goal,the ability to accurately model a user’s preferences from their historical interactions(for example,clicking,watching,reading,and buying)is key to an effective recommendation system.In order to model users’ preferences more accurately,social recommendation methods try to introduce users’ social information to achieve a more accurate recommendation system.However,some existing social recommendation systems still have some thorny problems.In view of these problems we tries to make the following improvements.First of all,due to personal privacy and trade secrets,the data sets of the recommendation system are often very sparse,and some users have less personal information and interaction records,making it difficult to model the preferences of these users.To solve the problem of data sparsity,we adopted the method of building collaborative user networks graph,compressed sparse user-item networks and user social networks,to enriched the relevant information of users with sparse interactive data.Secondly,the existing recommendation system often makes insufficient use of the information in the user’s social graph,and does not try to aggregate the information of higherorder neighbor nodes.We attempts to construct edges and static edge weights according to the social relations and interest similarity between users.We dynamically random walks on the dense collaborative user networks graph by edge weights to realize the sampling of high-order neighbors in the collaborative user networks graph.After obtaining the user set related to the target user,we further carried out aggregation propagation of the nodes in the sampling set through the graph convolutional networks,so as to complete the modeling of the target node.Finally,after introducing a variety of complex neighbor node sets based on interest,social relations,high order,or low order,it will contain valuable node information and noise irrelevant to the target node,so it is necessary to further remove the noise in the neighbor set and reduce the computational complexity.In this layer,we adopts the multi-head attention mechanism.Through the excellent performance and robustness of the attention mechanism,the sampled noise in the set is removed while the relationship between the neighbor node and the target node is learned.We verify the performance of the model on the Ciao and Epinions datasets,and compare our model with existing social recommendation methods.