Global Context Information Combined Graph Neural Networks For Session-based Recommendation

With the rapid development of science and technology in the new era,there is more and more data information on the network,which makes it more and more difficult for people to gain the data that they want in a limited time.For solving the problem of information overload,the existence of recommendation system is very important.However,in some cases,personal information or historical records may not be available or users’ interests may be changeable and immediate.In order to deal with such problems,session-based recommendation system has emerged.The purpose of the session-based recommendation system is to predict the users’ subsequent access behavior only according to the user’s current session information without accessing the user’s long-term profile.At present,session-based recommendation systems are mainly divided into three categories: traditional session recommendation methods,potential representation methods and deep neural network methods.In recent years,with the development of deep neural network,graph neural network has been widely used in session-based recommendation system,and achieved good results.However,most of the current session-based recommendation models have the problem of insufficient acquisition and use of the global relationship between items.For solving this problem,our paper combines the session-based graph neural network recommendation algorithm with the global context information,and proposes global context information combined graph neural networks for session-based recommendation,so as to make the most of the item transformation in all sessions to infer the user preference of the current session.This paper constructs the global graph and session graph according to all session sequences.In the global graph,the time interval of each interaction in the session,the second-order adjacent interaction information of the items in the session and the possible correlation between the adjacent interactions between the interaction items between different sessions are introduced to enhance the utilization of the global context information,which not only realizes the cross session information transmission,but also excavates the global potential information.The model will learn information from the session graph and the global graph respectively,and then obtain the item embedded representation by merging,and then use soft attention to obtain the session embedded representation to predict the next interaction behavior.In our paper,several experiments are executed on two benchmark data sets,which are compared with the existing session-based recommendation algorithms,and the model which is proposed in this paper is verified based on two recommendation evaluation indexes.In the final results,the model which is proposed in this paper has achieved the best results in various indicators,which shows the superiority of this model in the field of conversation recommendation.