Neural Collaborative Filtering For Social Recommendation Based On Graph Attention Network And Short-Term Dynamic Interest

Social recommendation is a research hotspot of recommendation algorithms in recent years.It can effectively solve the cold start and data sparse problems of traditional recommendation technology,and is widely used in e-commerce,movie and video recommendation,personalized advertising and other fields.Although the current social recommendation based on deep neural network has achieved good results,it still faces many problems.Learning accurate latent factor representations of users and items from user-item interaction graphs and social network graphs;capturing collaborative signals hidden in user-item interactions and describing the inherently complex interaction problems between users and items;furthermore,previous works usually ignore the correlation between items and rarely consider short-term dynamic representations of user interest and item attraction.Graph Neural Networks(GNNs)can naturally integrate node information and network topology,are powerful in learning node embedding representations,and can effectively utilize the high-order connectivity of graphs,which can improve social recommendation performance.Therefore,this paper studies the social recommendation method based on graph neural network.The specific work is as follows:(1)A neural collaborative filtering social recommendation AGNN-SR model based on graph attention network is designed.The AGNN-SR model is designed based on the basic social recommendation Graph Rec model and takes the user-item interaction graph and social network graph as input to improve the latent factor learning module and prediction module.In the latent factor learning module,users and items are modeled by adding a graph attention network(GAT),and the multi-head attention mechanism is used to learn more accurate latent factor representations of users and items from multiple perspectives.In the prediction module,neural collaborative filtering recommendation is designed by exploiting the high-order connectivity of graphs,which explicitly injects the collaborative signals of users and items into the embedding process to capture the deep interactions between users and items.The experimental results show that compared with the basic Graph Rec model,the MAE and RMSE of the AGNN-SR model on the Ciao,Epinions and Film Trust datasets are improved by 1.79%,3.78%,5.34% and 0.47%,2.31%,3.25%,respectively.(2)A neural collaborative filtering social recommendation DGNN-SR model based on short-term dynamic interest is designed.The DGNN-SR model is improved based on the basic social recommendation Basic DGNN-SR model.In addition to the user-item interaction graph and social network graph,the input module adds an item implicit network graph.In the latent factor learning module,for user interest and item attraction,the attention mechanism is used for static modeling,and LSTM is used for dynamic modeling.The final latent factor is obtained by combining the static and dynamic representations of users and items.The prediction module uses neural collaborative filtering recommendation in the AGNN-SR model to make predictions.The experimental results show that the MAE and RMSE of the DGNN-SR model on the Ciao,Epinions and Film Trust datasets are improved by 3.19%,3.88%,3.46% and 4.81%,3.55%,4.02%,respectively,compared with the basic Basic DGNN-SR model.Meanwhile,compared with AGNN-SR model without dynamic learning,the MAE and RMSE of DGNN-SR model are improved by 1.87%,1.51%,1.39% and 2.91%,1.41%,1.74% on Ciao,Epinions and Film Trust datasets,respectively.This paper also compares the AGNN-SR model and the DGNN-SR model with the comparison models such as PMF,So Reg,Social MF,etc.The experimental results show that the recommendation effect of the models in this paper is better than that of the comparison models.