Research On Recommendation Algorithm Based On Graph Convolution Neural Network

With the rapid development of the Internet,people are indulged in a huge amount of information and find it difficult to make effective decisions quickly.The emergence of recommendation algorithms has improved the problem of information overload by filtering a large amount of redundant information for users,and gradually become an indispensable tool in life.The main idea of collaborative filtering algorithm is to model the interaction between users and the subject matter,and it has been a widely used recommendation algorithm from the beginning of its birth until now,and the research on collaborative filtering algorithm has never stopped over the years.In recent years,as the research on graph structure deepens,the collaborative filtering model based on graph neural network starts to receive widespread attention.It has two advantages: first,the collaborative filtering data itself is a graph structure,and the graph neural network used to process the graph data fits well with it;second,the prediction problem of recommendation can be regarded as the link prediction problem in the graph structure,and the graph neural network is born to solve such a problem.However,current collaborative filtering methods based on graph neural networks still have shortcomings: first,the model ignores the effect of layer combination on the final node embedding representation;second,the model has problems in handling newly added users and items.Therefore,the following research work is carried out in this paper.First,the current collaborative filtering algorithm LightGCN based on graph convolutional neural network that can achieve the best performance is analyzed in depth,and the experiments are compared by adding layer combinations and directly using the last layer as the final embedding representation of the nodes respectively,and it is demonstrated that the importance of layer combinations for the final node embedding representation and the embedding representation obtained from each layer of graph convolution has different importance for the final embedding representation The attention mechanism is therefore introduced into the layer combination.Then a series of comparison experiments are conducted on various datasets,including performance comparison with different layers,performance comparison before and after improvement,comparison with frontier algorithms,and exploration of some hyperparameters,which demonstrate that the improved algorithm with the addition of attention can effectively improve the recommendation performance.Second,to address the problem that most current collaborative filtering algorithms based on graph convolutional neural networks cannot handle new users or new project nodes,the UOEGCN framework is proposed,which can achieve inductive inference and only requires one of the user or project embeddings for training,significantly reducing the number of model parameters and improving the training speed,solving the problem that existing models are complex and difficult to train and scale It solves the problem that existing models are complex and difficult to train and scale.Finally,the framework proposed in this paper is combined with LightGCN to perform experiments on the UOELGCN-U model using the initial user embedding representation to verify the effectiveness of the model and compare it with LightGCN and NGCF to significantly improve the training speed and reduce the number of model parameters at the expense of a certain accuracy rate.The advantages and disadvantages of the UOELGCN-U model and the UOELGCN-E model using the initial item embedding representation are also experimentally compared,and the analysis yields that the UOELGCN-U model is more suitable in practical applications.