Research On Social Recommendation Model Based On Graph Neural Network

The recommendation system has always been one of the key technologies to solve the problem of modern information overload.In the development process of the personalized recommendation system,it has been faced with problems such as data sparsity,cold start,and poor performance of recommendation system,which have a great impact on the accuracy of recommendation system.In order to solve this problem,researchers have introduced social relationship attribute information into the traditional recommendation framework and proposed a social recommendation methods.At this stage,most methods only utilize first-order social relationships,and do not fully utilize the auxiliary information in the social graphs,while higher-order social relationships can better explain users' interests and preferences through friends with different distances.To address this problem,this thesis uses two different graph neural network methods to capture high-order neighbor representations in users' social relations,thereby improving the performance of recommender systems.The main contents of this thesis are as follows:(1)In order to effectively extract the impact of social neighbors with different distances on users' preferences,this thesis proposes a social recommendation model integrating multiple strategies.The model generates users' social sequences by random walking on the social graph,and uses the two-way long-term and shortterm memory network with attention mechanism to capture the interaction opinions of neighbors on the project when users interact with the project,so as to obtain the user potential vector.Use the graph to pay attention to the potential vector of elearning items,and finally predict the score in the neural collaborative filtering module.Based on two real-world datasets,the model is compared with the representative and advanced models in the field of social recommendation.The influence of the length and number of sequences on recommendation performance is analyzed,and the validity of the model is proved.(2)In order to solve the problem of increased computation due to the increase of high-order neighbors,this thesis proposes a social recommendation model based on a multi-layer graph neural network,which uses a multi-layer graph convolutional network to recursively embed along the social network.Propagation,effectively capturing the influence of high-order neighbors on user preferences,and using an attention mechanism to aggregate friends from different layers,resulting in better vector representations.A graph attention network is used to obtain potential representations of user and item perceptions,and finally rating predictions are made in a neural collaborative filtering module.The model is compared with the representative and advanced models in the field of social recommendation based on two real-world datasets,and the impact of the embedding size and different layer aggregation methods on the recommendation performance is analyzed,which proves the model has better performance in capturing high-order neighbor representation.