Research On Multi-interest Recommendation System Based On Graph Neural Networks

With the vigorous development of artificial intelligence,intelligent products have entered every household and gradually penetrated into our work,study,rest and entertainment,social interaction,industrial production and other fields.With its intelligent and personalized features,the Internet has provided great convenience for human beings,and the means of human communication have become rich and diversified.At the same time,with the explosion of information and data on the Internet,our daily life is filled with all kinds of data,which makes what we really need and care about disappear in the ocean of data,resulting in a huge phenomenon of information overload.Recommendation technology has surpassed the early search engine as an indispensable means of data screening,which has greatly promoted the development of e-commerce and other fields.Therefore,how to improve the performance of recommendation system has become the current research hot pot.In recent years,Graph Neural Networks has shown great superiority in processing graph structure data,and has been widely used in the field of recommendation system.The depth mining of potential information through graph neural network is beneficial to improve the accuracy of recommendation.However,the recommendation system based on graph neural network still has some shortcomings.In the traditional collaborative filtering recommendation model,due to the lack of interactive information of the target node,the final representation of the node is not accurate,that is,the data sparse problem.At the same time,most of the models only consider the continuous acquisition of deep-level information,but ignore the two sides of information,and the aggregation of high-level information without distinction will eventually lead to the deviation of recommendation results.To solve these problems,this paper proposes a multi-interest recommendation model based on user clustering to perform recommendation tasks.In addition,in the process of increasing the number of layers,the model implemented by the graph neural network technology will appear too smooth,which limits the performance of the recommendation system to a certain extent.And the traditional recommendation model focuses too much on the accuracy of recommendation and ignores the multi-interest of users.This thesis further proposes a multi-interest recommendation model based on neighborhood interaction.The research content and innovative work of this paper are as follows:(1)The profoundly research into the theory about multi-interest recommendation based on graph neural network.This thesis first gives a description of the involved theories about recommendation system,then concludes some challenges faced by traditional recommendation algorithm research and studies the existing recommendation models based on graph neural network,and analyzes the shortcomings of them.Based on this,this thesis puts forward the research of multi-interest recommendation system based on graph neural network,and describes the key technologies used in the research.(2)Multi-interest recommendation model based on user clustering is proposed.(CMI-GCN)With regard to the problem of sparse data,one of the factors affecting the poor recommendation effect in the current traditional collaborative filtering recommendation system,user clustering is integrated into the recommendation system.For the sake of working out the problem of sparse data and the problem that the user’s feature representation is difficult to be accurately learned because user information cannot be deeply mined,this thesis proposes a multi-interest recommendation model(CMI-GCN)based on user clustering.Before collaborative filtering and recommendation,the model firstly uses clustering method to cluster users,selects information retrieval space for feature learning of target users,and then uses graph convolutional neural network to deeply explore the global influence on feature learning of target users.The comparison experiment proves that the CMI-GCN model can effectively alleviate the problems existing in the traditional recommendation algorithm,and significantly improve the recommendation performance.(3)Multi-interest recommendation model based on neighborhood interaction is proposed.(MI-GCN)Most current models based on graph convolution neural network achieve the best performance by adding convolution layers.However,in the process of continuously increasing the number of layers,the feature representation of each node will gradually reach approximately the same state,resulting in inaccurate recommendation results.At the same time,direct reference of higher-order information will produce higher-order noise,which will lead to the deviation of node representation learning.Therefore,this paper proposes a multi-interest recommendation model(MI-GCN)based on neighborhood interaction.Subgraph generation mechanism is designed in this model,users are divided by three different methods,user characteristics and their interactive items are combined with user-project bipartite graph to generate subgraphs,so that users with multi-interest can belong to different subgraphs,and more accurate representation of user characteristics can be obtained by aggregation of information in subgraphs.Experimental results show that the MI-GCN model can effectively alleviate the over-smooth problem and high order noise problem,so as to improve the accuracy of recommendation results.