Research On An Efficient Non-sampling Graph Neural Network Model For Social Recommendation

At present,the Internet has become a technical field with widespread participation and far-reaching influence.While giving full play to the advantages of informatization to bring benefits and convenient services to the people,it also faces a series of challenges such as information overload.Recommendation system helps active users to obtain information that meets their needs through information filtering,which plays an indispensable role in daily rapid decision-making process.As a hot research direction for many researchers,social recommendation integrates social relations on traditional recommendation methods by utilizing user-item interaction record to model user preferences,which has improvements in recommendation performance.In recent years,graph neural network has been developed to improve the representation learning process of users by aggregating high-order neighbor nodes through simulating the recursive diffusion process of social influence.However,most of social recommendation methods based on graph neural network have two key problems that have not been well studied:(1)from the perspective of model design,they are usually in the public space of the user characteristics attribute mapping for low dimensional potential said,ignoring the different users in the field of social and information inherent differences between user items field;Furthermore,the high-order connectivity between nodes reflected in the social graph and interaction graph is not taken into account at the same time,and hidden feature information is mined for better user/item representation.(2)In the aspect of model optimization,most recommendation methods rely on negative sampling to optimize the model,and recommendation performance will be influenced by the quality and quantity of negative sampling,which makes some models highly sensitive to the design of the optimizer and difficult to make full use of the computing power of GPU.In this paper,we propose an efficient non-sampling graph neural network(EAGNN)for social recommendation,the main contribution points are as follows:(1)considering the difference of behavioral information richness of users in social domain and interest domain,fine-grained learning of the feature representation of user nodes based on specific relationships;On the other hand,bilinear graph convolution module is designed to jointly simulate the recursive diffusion process of social influence on social graph and interest synergy effect on interaction graph.High-order neighbor information is aggregated to target users.In this way,more information beneficial to user and item embedding learning is mined.(2)The user feature adaptive fusion method based on the gated unit mechanism is adopted,and the adaptive learning social influence contributes to the user’s final embedding.In this way,the EAGNN model can judge the importance of social influence propagation in predicting different user preferences.(3)To ensure the training efficiency of the model,an efficient optimization algorithm based on non-sampling strategy is designed to update learning model parameters through matrix operation with lower complexity,which improves the GPU utilization rate and greatly improves model training rate.A large number of comparative experiments were performed on four datasets to evaluate the EAGNN model.The results show that EAGNN model can show significant advantages in performance and training efficiency in different recommendation scenarios,compared with state-of-the-art recommendation methods.It verifies that graph convolutional networks and non-sampling strategies have effectiveness on social recommendation.