Application Of Graph Neural Networks In Recommendation Systems

Recommendation systems filter out a large amount of useless information for users and help users make effective decisions,which has become an indispensable tool in life.As a classic recommendation algorithm,collaborative filtering algorithm models the interaction between users and commodities based on implicit or explicit feedback,and is the most widely used and effective recommendation algorithm.Therefore,how to improve the recommendation quality of collaborative filtering algorithms has become a hot issue in algorithm science research.In recent years,scholars at home and abroad have proposed many model frame-works to improve the quality of collaborative filtering recommendations.The collabo-rative filtering model based on graph neural network architecture is an important part of it.The data of collaborative filtering is essentially a graph structure,and the method of constructing a graph to model the collaborative filtering problem conforms to the law of the data itself;and the recommendation prediction problem is mapped to the link prediction problem in the graph structure,and the graph neural network is used in the link prediction problem and The graph data shows that it has a unique advantage in learning.Therefore,the collaborative filtering model based on graph neural network can improve recommendation quality and optimize recommendation efficiency.How-ever,the current collaborative filtering method based on graph neural network still has shortcomings:first,the processing paradigm of the current graph neural network model is to expand the original interaction graph in a tree,which loses the location topology information of the interaction graph;second,the model pair Insufficient attention to popular items leads to more frequent recommendations of popular items,which affects user experience.Based on the above analysis,this paper designs two collaborative filtering models based on graph neural networks.The main research work and innovation are as follows:1.This article points out that the training paradigm of the current graph neural network is essentially a tree-like expansion of the original interaction graph with the target node as the root node.This training paradigm will destroy the topological location information of the original interactive bipartite graph.To solve this problem,this paper proposes a graph collaborative filtering model with topological position coding.The model uses the METIS graph segmentation algorithm to capture the topological position information of the interactive graph and form a position code to improve the quality of collaborative filtering embedded representation.At the same time,the model simplifies the graph convolutional network module in the current graph recommendation system,and improves the recommendation efficiency.2.This article focuses on the imbalance in the recommendation system of the cold and popular items in the recommendation system.In this paper,the pre-trained graph collaborative filtering model is used to simulate the user's instant feedback,and then a graph convolution collaborative filtering model based on Bandit enhancement is constructed.The model uses the Bandit algorithm to capture the user's potential purchase interest by constructing a virtual graph,and makes an attempt for users to explore interest,which increases the probability of the model recommending unpopular items,thereby alleviating the imbalance of popular and popular recommendations.In order to verify the effectiveness of our proposed model,this article conducted corresponding experiments on three real recommendation data sets Gowalla,Yelp2018,and Amazon-Book.Experiments prove that the introduction of position coding has a significant impact on the recommendation results.Compared with the baseline method,the graph collaborative filtering model with position coding improves the Recall and Ndcg indicators by up to about 7%.At the same time,multiple experiments have proved that the graph collaborative filtering model based on Bandit enhances the model's ability to learn unpopular items,and alleviates the problem of unbalanced recommendation of popular items.