Research On Graph-based Collaborative Filtering Recommendation Algorithm

The information filtering technology,with recommendation system as the core,is an effective means to alleviate the problem of information overload.Nowadays,personalized recommendation technology is one of the key technologies to improve user experience.Recommendation algorithm is the core part of the recommendation systems.Collaborative filtering algorithm occupies an important position in the field of recommendation systems,which is widely used and mature enough in the industry,and has the advantages of simplicity and efficiency.In addition,the recommended data is large and sparse.Due to various factors,there is noise in the recommendation data,which cannot be extracted comprehensively and efficiently.Most of the data in the recommendation system is complex non-Euclidean structure data,such as bipartite graph is composed of user-item interaction and the social network in the social recommendation.In recent years,graph learning methods,especially graph neural network and attention mechanism,have been widely used in the field of recommendation systems,which has inherent advantages in processing graph-structured data.The collaborative filtering algorithm based on the graph neural network can learn the deep representation between users and items.It regards the recommendation problem as the connection between nodes and message propagation,and carries out modeling to realize the feature extraction of graph structure,so as to achieve the purpose of recommendation.The key of collaborative filtering recommendation algorithm is to use an effective function to calculate the similarity between users or items.The similarity function has a direct impact on the performance of the model-based method,which is conducive to improving the accuracy of the model.This thesis is dedicated to the study of collaborative filtering algorithm based on graph structure.It focuses on improving the feature extraction method of recommendation data,combining similarity and graph neural network to improve recommendation performance.This thesis mainly studied and discussed from two aspects:the local graph attention collaborative filtering recommendation method and the similarity-based graph attention collaborative filtering recommendation method.The research work and contributions are summarized as follows.First,a local graph attention collaborative filtering recommendation method is proposed.This method improves the data utilization mode.On the one hand,the shallow linear structure of user-item interaction data is extracted by singular value decomposition,and the noise in the data is filtered.On the other hand,the noise-filtered data is passed through a graph attention network to learn embedding representations for extracting deep nonlinear structures that cannot be extracted by traditional collaborative filtering algorithms.The data after fully feature extractions is propagated through graph convolution module.Finally,the inner product is used to estimate the user’s preference for the target product.Experimental results show that the proposed method exhibits good recommendation performance.Second,a similarity-based graph attention collaborative filtering recommendation method is proposed.This method uses the user-item interaction data to calculate the similarity-based embedding representation for user and item through the Gaussian interaction distribution kernel function,and realizes the aggregation and dissemination of user and item information based on multi-layer lightweight graph convolution propagation rules.Then,the high-order feature embedding representations of users and items are extracted,and the attention mechanism is used to combine the high-order feature embedding representations to realize the modeling of users and items with high-order structural information.This method fully combines the similarity of traditional collaborative filtering with the graph structure information connectivity of collaborative filtering algorithm based on graph neural network to model users and items.The experimental results show that the model can effectively improve the recommendation performance to a certain extent.