Research On Package Recommendation Based On Graph Attention Network

At present,with the rapid development of the mobile Internet and the rapid popularization of smart terminals,people are already in an era of information overload and explosion.Information producers continue to output information,and information consumers continue to receive information,which creates some problems.It is difficult for information producers to show information consumers what they want,and it is also difficult for information consumers to find the information they want from a large amount of information.To solve this difficult problem,technicians have developed recommendation systems.Since then,recommendation systems have developed rapidly,building a bridge between users and information.The recommendation system can help users find the information they are interested in,and can also display the information to relevant users,so as to realize the information matching between information producers and consumers.Recommendation systems have been widely developed and applied to various websites and platforms to promote products or services to users in a targeted manner.In quite a few sales scenarios,the sales platform needs to expose a series of items to users,which is package recommendation.However,there are few researches in this area and there are still many deficiencies.This paper develops a novel and practical package recommendation system PGAT(short for Package Graph Attention Network)based on graph neural network.PGAT integrates users,items,and packages to build a unified heterogeneous graph and treat them as a whole.PGAT adds an attention mechanism to the first-order neighborhood aggregation operation,which can learn the weights of different neighbor nodes to the central node.By performing graph attention and graph convolution operations on tripartite graphs,PGAT can learn node embedding more efficiently and solve the problem of data sparsity to a large extent.Extensive experimental results on two real-world datasets validate the excellent performance of PGAT,outperforming the state-of-the-art baselines by 0.77% to 10.12%.The research content and innovation work of this paper are as follows:(1)Study the principle of recommendation system and package recommendation deeply.This paper first introduces the research overview of recommendation systems in the era of machine learning and deep learning,and then introduces the research status of package recommendation systems.In order to solve the problems existing in the current meal recommendation algorithm,a research plan of meal recommendation based on graph neural network is proposed,and the main related algorithm technologies are introduced.(2)Propose a graph neural network package recommendation model(PGCN)based on three-part graph.On the basis of user-package interaction,the interaction between users and items,the affiliation between packages and items,and the similarity between packages are introduced to construct a tripartite graph of users,items,and packages.Introducing item nodes and using them as bridge nodes can introduce additional information of user interests and further alleviate the cold start problem of user package recommendation.Unifying the heterogeneous graph composed of user nodes,item nodes,and package nodes into a whole,the model can fully mine and extract the connections between them for package recommendation.(3)Propose a graph neural network package recommendation model(PGAT)based on the attention mechanism.Drawing on recent advances in geometric learning,this model introduces graph attention networks and graph convolution networks to model complex relationships among users,items,and packages to solve the package recommendation problem.We develop two different embedding propagation layers,a multi-head graph attention layer and a graph convolution layer.The multi-head graph attention layer combines an attention mechanism to distinguish the impact of different items in the package on user decision-making,which can improve the expressive performance of the embedding.Due to the over-smoothing problem,it is difficult for multi-head graph attention layers to aggregate high-order neighborhood nodes and distinguish the importance of different nodes in high-order neighborhoods.We use graph convolution layers to compensate for its shortcomings.(4)Experimental analysis,and employ T test to verify performance improvement.Extensive experiments on two real-world datasets outperform existing state-of-the-art baselines by 0.77% to 10.12%,verifying the effectiveness and advancement of our proposed model.In addition,we performed a T-test on the experimental results to show the improved significance level.