Research Of Recommendation Method Based On Graph Neural Network

With the rapid development and high popularity of the Internet in China,information overloading makes it a serious challenge for crowds to access the content they are excited about.The rise of recommendation systems,which not only improve user stickiness and loyalty,but also learn about users’ potential interests and preferences through information on user attributes and their historical interactions,can help save users’ time in sifting through information,improve their shopping experience and alleviate information overload during the shopping process.In general,classical traditional recommendation algorithms only consider the historical interaction behavior of users and goods,and tend to ignore the intricate graph structure of users and goods that exists in real recommendation scenarios,as well as the problem of merchants recommending multiple goods to users in real recommendation scenarios.In contrast,graph neural network can aggregate information about neighboring nodes directly on the graph network through graph convolution operations,which can be applied in non-Euclidean structured domains to effectively capture higher order representation relationships between data,and are currently used in many recommendation scenarios.In this context,we propose two different graph neural network-based recommendation algorithms that combine graph neural network to learn complex representation relationships between users and goods for recommendation.The main work of ours has two main points.Firstly,most of the classical traditional goods recommendation algorithms have difficulty in capturing the higher-order information of interactions,while graph neural network can catch the complicated topological structure information and retain the graph structure information of attributes in the graph network,to extract richer structural representation information and semantic representation information and improve the goods recommendation effect.Therefore,we proposed a multi-modal personalized product recommendation model based on Graph Attention-Enhanced Graph Neural Network by combining graph neural network related methods and using user and goods interaction behavior data.The model makes full use of user and goods interaction behavior data as well as user and goods related textual information,which helps to alleviate certain data sparsity and overfitting problems.On several datasets,the proposed model outperforms some comparable benchmark models and obtains a relatively good recommendation.Secondly,most of the previous research on goods recommendation algorithms is to recommend individual goods to users,however,in the real recommendation scenario,the platform needs to be recommending a collection of goods to users,for example,the marketing strategy will package multiple goods for sale together,which is called bundled recommendation in the real recommendation scenario.The bundle recommendation can solve the problem of difficulty in choosing multiple goods for some users,which is a great help to improve the shopping experience of users.Therefore,we proposed a Bundle Recommendation based on Graph Attention,which learns complex node representations of heterogeneous associations from a user-item-bundle heterogeneous network,from review text extracting textual semantic information,learning social semantic information representations from users’ social networks that affect users’ bundle recommendations,and then aggregating these rich semantic embedding through an aggregator for alleviating certain data sparsity problems.Finally,we performed extensive experimental comparisons using publicly available datasets.The experimental results show that we proposed method significantly improves the efficiency of the recommendations compared to the reference method.