Research On Graph Neural Network Recommendation Algorithm Based On Global Information Session Sequence

With the high-speed development of computer technology and e-business,a variety of shopping goods,books,news articles,movies and documents present a trend of exponential explosion.Recommending systems,which are efficient solutions to information overload,not only assist users to make rapid choice in the ocean of information,but also help online platform increase user conversion rate and economic benefits.Moreover,session-based recommendation gains more popularity for more accurate predictions from anonymous sessions.Recently,graph structure got high attention due to its natural adaptability for sessions,thus many researchers dived into graph neural network based recommending algorithms and achieved state-of-the-art performances.Although these methods could make relatively accurate recommendations utilizing graph structural information,they neither consider repetitive submissions from users and complex transitions between items,nor make full use of complex graph structural information.Worse still,neural networks without enrichded information propagation eventually reach a state where over-smoothing exists and consequently they all suffer from prediction loss to a certain extent.Aiming at the shortcomings of the existing research work,this paper mainly makes the optimization improvement from the following aspects:(1)Based on the original model,the initial eigenvectors are processed,and global information is extracted more comprehensively by using dynamic global neighborhood attention and multi-channel graph neural network to create session diagram,global diagram and consensus diagram flexibly.(2)This paper introduces a graphical location encoder to embed the consensus graph which combines session diagram and global diagram,and obtains session representation with position embedding,so as to obtain more accurate global information,lay the foundation for extracting global information for subsequent operations,and optimize input data of the original model.(3)Propose a new wandering strategy with attention.On the basis of the inert random wandering strategy,the propagating nodes themselves have sufficient attention weight,so in the process of feature propagation,the graph nodes can balance the ratio of adjacent node information to their own information,and avoid the situation that the graph information enters the excessive smoothness too early and gets the bad recommendation effect.This paper conducts an experiment using three datasets,and the experimental results verify that the methods and models proposed in this paper are superior to some existing methods and more accurate in predicting users’ next click.