Research And Application Of Collaborative Filtering Recommendation Method Based On Graph Convolutional Network

With the rapid development of Internet technology,the production and transmission of information has far exceeded the speed of people’s acceptance,and the phenomenon of "information overload" has become more and more significant.Recommendation systems can effectively reduce information overload by capturing users’ interest points and filtering personalized information from a large amount of information.Graph convolutional networks are widely used in recommendation systems because of their ability to fuse node information and topology with graph structure data as input.The problems of existing recommendation models based on graph convolutional networks include coarse granularity of user behavior modeling,inability to effectively distinguish the importance of neighboring nodes in the process of node information aggregation,insufficient use of node history information in node information updating,and over-smoothing when stacking too many network layers.In this paper,the research and application of recommendation methods based on graph convolutional networks are carried out to address the above problems.The main work and contributions of this paper are as follows:(1)In this paper,the research is based on the idea of collaborative filtering,using the historical user-item interaction data to implement the Top-N recommendation task.The study uses user-project bipartite graph data and proposes an enhanced graph convolutional network collaborative filtering recommendation model combining graph attention(AEGCN).In the neighborhood aggregation part,the neighbor node features are aggregated using an embedded propagation layer combining graph attention and graph convolution,and the user preferences are fully learned through propagation on the graph to effectively distinguish the importance of neighbor nodes;in the node update part,the self-information enhanced node update algorithm and add node history information as input to enhance feature learning.To verify the performance of the proposed recommendation model,comparison experiments are set up on three public datasets,and the results show that both performance indicators of the proposed AEGCN model outperform other benchmark algorithms in terms of recall and NDCG,and the performance and effectiveness of the model are verified by setting up ablation experiments.(2)This paper proposes a residual enhanced graph convolutional network collaborative filtering recommendation model(R-AEGCN)for the over-smoothing phenomenon that may occur when the number of network layers is stacked too much.The model adds a residual structure module to the AEGCN model,which can effectively improve the recommendation effect of the model by extracting the residual preferences of users while effectively alleviating the oversmoothing phenomenon.After experiments on three publicly available datasets,the results show that the R-AEGCN model further improves the recommendation performance compared with the AEGCN model on the basis of two performance indicators,namely,recall and NDCG,which are both better than the benchmark model.At the same time,it is demonstrated that this model can effectively alleviate the over-smoothing problem caused by too many layers stacked in the graph convolutional network by setting the hyperparameters of network layers.(3)In the movie recommendation scenario,a personalized movie recommendation system is designed and implemented on the basis of the proposed model R-AEGCN in this paper.The system adopts the layered architecture of MVC pattern,and the front-end part adopts Vue framework and Element UI component library to realize the collaborative filtering recommendation task based on the user’s historical interaction data.The system can provide users with a movie library,and users can rate and collect movies in the library.The system learns users’ preference features and generates personalized recommendation lists for users after model training.The system can effectively filter the movie information of interest to the user and alleviate the information overload problem.The model can be transferred to other recommendation scenarios.