Research On GNN Link Completion Algorithm And Its Application In Film Recommendation

In recent years,the graph neural network based recommendation system has shown excellent performance and prospects,and many excellent results models have appeared,but there are still many problems to be solved: 1.Typically,the criteria for a user to rate items are similar,so items that receive the same rating have similar properties,but most model sampling methods ignore this underlying relationship.2.The aggregation of node information pays too much attention to the neighboring nodes and does not pay attention to the structural information of distant nodes.3.In the process of node aggregation,user node and item node are regarded as the same kind of node,which affects the performance of graph neural network.4.It is not noticed that the influence of each node is different and should be treated differently among nodes of the same type.In view of the above problems,this thesis completes the following work:1.This thesis presents a graph neural network algorithm based on hierarchical sampling,through the link between completion node thus recommendation algorithm,algorithm include:(1)a hierarchical sampling method based on the diagram,the method will be carried out in accordance with the interactive data graph partition,make the same type single polymerization between nodes,envoys point features more evident,to solve the problem of node similarity.(2)Unit independent training is adopted to train each subgraph separately.Finally,the output of subgraph training is classified and summarized to output link prediction and score prediction respectively,so as to solve the same problems of user node and item node.The effectiveness of the method was verified in the self-comparison experiment,and the most suitable model was selected to compare with the baseline model.The RMSE,m AP@10 and n DCG@10 indexes of the best training model of the algorithm were improved in the selected data set.2.A hierarchical link prediction method based on graph attention mechanism is proposed,which uses the graph neural network framework of hierarchical sampling and includes(1): Node2 Vec method of first-order breadth search priority algorithm is used for node embedding,which makes node embedding pay attention to the characteristics of neighboring nodes,and uses cross-neighbor node aggregation in the process of aggregation to solve the problem that distant nodes are not paid attention to in the process of aggregation.(2)Graph attention mechanism is adopted to distinguish different influences of nodes on aggregation,and more influential nodes are allowed to participate in link prediction and scoring prediction,so as to distinguish problems with different influences of different nodes.Furthermore,the graph neural network model with higher level sampling on the selected five data sets is further improved.3.Finally,Spring Boot and u Wsgi services are adopted to apply the multi-layered link completion model based on graph attention mechanism to the recommendation system to realize movie recommendation,and Mysql database is used to realize data persistence,and Redis database is used to realize movie evaluation ranking within a week.And through engineering means,generate the recommendation list according to the user preference and expand the user interest point recommendation list to alleviate the problem of repeated recommendation and "information cocoon house".