Research On Recommendation Algorithm Based On Graph Neural Network

Recommendation systems play an increasingly important role in improving the user experience of network products.The essence of a recommendation system is to portray the user and predict what items the user may interact with.Although the recommended model based on collaborative filtering has always been the basic paradigm,the key to improving model performance is to model collaboration signals between users and items.Because the recommended algorithm of graph neural network has advantages in capturing collaboration signals between users and items,the combination of graph neural network and recommendation task has become a research hotspot.Previous studies have found that there are some problems in coding node characteristics of Graph Neural Network,and these problems are studied in this thesis.When the graph neural network is combined with the recommended task,on the one hand,there will be the problem of useless links in building the interaction graph due to noise in the historical interactive data.On the other hand,the importance of recognizing neighbors based on the graph structure when the graph convolution neural network is embedded in the propagation impedes the representation learning of the graph nodes.For this reason,a recommendation model based on graph-based contrastive learning and attention mechanism is proposed.In the model,three graph data enhancement strategies,random exit of graph nodes,random loss of connection edges and random walk,are used to reduce the influence of noise data on the propagation of graph embedding.The model fuses attention mechanisms at the graphical convolution neural network layer to dynamically assign weights to neighbors of different nodes,which improves the node feature representation ability of the graphical convolution neural network.The recommended model based on a convolution neural network can cause excessive smoothing of node characteristics when overlaying too many graph convolution layers.In addition,the embedded propagation based on the bipartite graph also makes the model unable to learn the global synergistic signals of user nodes and project nodes.A hypergraph neural network recommendation model based on gated mechanism is presented.This model adds hypergraph embedding propagation based on ordinary graph embedding propagation,which enables nodes to obtain both local and global collaboration signals.he gating mechanism in this model can improve the iteration update mechanism of graph node characteristics,which can suppress the excessive smoothing effect of nodes and keep more effective feature components in the iteration update.The two models proposed in this article have verified the improvement in recommendation performance on publicly available datasets,and the effectiveness of each module of the model has been verified through ablation experiments.