Research On Graph Neural Recommendation Model Based On Item Sequence Dependency

In the age of Big Data with the rapid development of the Internet,recommendation algorithm is the key method to solve the “Information Overload”,therefore,it has become a research hotspot in academia and industry.How to mining user’s interesting and item sequence dependency from high-dimensional sparse user historical browsing data to solve data sparsity is the most significant research problem that need to be solved urgently in the community of recommender system.At present,most recommendation models based on deep learning methods uses additional auxiliary information to solve this problem,it causes additional cost of data acquisition and these methods doesn’t adapt to the graph structure data.Therefore,this dissertation based on the graph neural network,deeply mining user-item historical interaction data,proposes two new recommendation models to strengthen the expression of user and item.(1)This study proposes an item sequence dependency recommendation model based on graph neural network.The main purpose is to mining item implicit relationship by using item dependency to alleviate the problem of data sparsity.Firstly,this study construct item sequence dependency relationships graph by user historical browsing data to reduce the time to abtain additional auxiliary information.Secondly,this study captures the item dependency by using the classical graph convolution neural network to strengthen the expression of item.Then,this study using graph collaborative filtering method to embed the user-item interaction and non-linear relationship in user and item embedding vector,to further enhance the deeper representation of user and item.Finally,this study exploits loss function training model continuously to improve recommendation performance.(2)This study proposes an item sequence dependency recommendation model based on graph attention network.The main idea is to accurate the expression of user and item by quantifing the correlation of items to capture the attribute characteristics in a more granular way.Firstly,this study uses Multi-layers perceptron to map the higher-dimensional information into the low-dimensional space.Secondly,this study construct item sequence dependency relationship graph by sliding window strategy to use graph attention network evaluating the item relevance from multiple dimensions.Finally,the ability of graph convolution to extract data features in graph structure can improve the expression of users and items.The two recommendation models based on graph neural network are tested on three public data sets(Last FM,Douban,Caio).Compared with other classical recommendation models,the experimental results show that ours’ models are superior to other models in terms of performance.Next,this study analyzes the two models of the time and space complexity and analyzing the important hyperparameters in tow models,the experimental results prove that the two models can improve the recommendation performance and stability to a certain extent in the face of data sparsity.