Research On Deep Recommendation Algorithm Integrating Social Networks

Traditional recommendation algorithms model users' preferences for items by analyzing the interaction data between users and items.Limited by the sparse interaction data,such models may not be able to learn enough information to accurately model user preferences.In order to cope with this problem,scholars tried to add other information related to users and items into the recommendation algorithm as a supplement of interaction data,social network recommendation is a kind of method adding social information of users to enrich sparse interaction data.Most of current social networks recommendation algorithms only utilize direct friend relationship between users(first-order connections).These methods have limitations,because there is also sparsity problem in user social data.In addition,the distribution of social data is uneven.Users with very rich social relationships and users with very sparse social relationships all occupy a large proportion among all users.This kind of distribution makes social data substantially more sparse.When the first-order social information among users is sparse,these first-order social models may not achieve satisfactory results Therefore,it is not enough to use only the first-order connections in social relation data.Social relationships can be represented as graph-structured social network,GNN techniques in deep learning can easily exploit the higher-order information in graph structure through neighbor aggregation.This provides a new method to model higher-order social relation.Based on these assumptions,this paper have done the following work:(1)This paper proposes two social networks recommendation models on graph structures,they can not only organize user social ralation into graph form to exploit higher-order information between users,but also use modified cosine similarity to analyze the common rating data of different users on items,discover the similar connections between items and organize them into graph form.The item ralation graph constructed by this method not only contains the contact information and high-level contact information between items,it is also an important manifestation of high-level interaction information between users and items.By utilizing these high-order contact informations,the data sparse problem faced by the algorithm can be greatly alleviated.(2)The user-item interaction data contains not only rich information but also a lot of noise information irrelevant to the current task.In order to make better use of these data,this paper designs an attention model to extract the most relevant information from the user-item interaction data.(3)In order to explore high-order connections between users and items,this paper applys graph convolutional neural networks on social networks graph and item ralation graph to learn the higher-order feature representation.of users and items.(4)To solve the problem that the classical graph convolutional network cannot flexibly learns the influence weights between nodes,this paper uses the graph attention network to improve the model.Besides,in order to adapt to the uneven distribution of connection relationships in social relationship data and item relationship data,attention model is designed to learn different feature combination schemes for different types of nodes,in this way,the model can flexibly integrate node' representations of each hop.(5)Comparing experiments with current classical recommendation algorithms are conducted on two open source datasets,the experimental results demonstrate the effectiveness of models proposed in this paper.