Research On Session Recommendation Model Based On Graph Neural Network

In recent years,the rapid development of mobile Internet and the popularity of intelligent terminals have led to a blowout growth in the type and quantity of online resources,which has brought convenience to users and also brought problems such as information overload.In order to overcome the severe problem of information overload,we need to design a recommendation algorithm to process a large number of interactive information between users and goods,and screen out goods that are more in line with users’ interests and preferences to recommend to users.The conversational Recommendation system studied in this paper is an important branch of the Recommendation system.According to the user’s item click sequence in a certain period of time,it can predict the user’s next click and help the user make decisions.Through learning and comparing the traditional machine learning based session recommendation systems,as well as the currently popular neural network based recommendation systems,we found that the current session recommendation systems have the following problems: First,most session recommendation systems only focus on the information interaction within the session,without taking into account the complex project conversion relationship between sessions;Second,the model has limitations in learning complex representations,and cannot be converted quickly and accurately;Third,most systems only focus on static settings and batch process data,ignoring the streaming nature of session data.In order to solve the above problems,this paper proposes two improved global and local graphical neural networks(IGL-GNN)and graphical neural networks with user information(UGNN)models.The specific work is as follows:First,the improved global and local graphical neural network can not only model the project transformation in the target session,but also model the project transformation between sessions.Specifically,we learn the local level item representation from the session graph and the global level item representation from the global graph.Then we explored the aggregation method CNN pooling to more effectively aggregate global and local information,which not only retained more useful information,but also deleted useless information.In addition,we innovative introduced Transformer and included it in the model as a general deformation function,which enhanced the ability to obtain complex transformations and solved the limitations of the model in learning complex representations.Second,the UGNN model with user information.When building the graph structure,we introduced general user configuration information to facilitate the update of the model and more personalized recommendations.The Wasserstein repository part of the model makes the session stream data no longer batch processed,but can be continuously and dynamically updated like a pipeline to constantly capture the latest session information.The system in the streaming environment is not static,but dynamic and updated in real time,which can adapt to the high-speed and continuous arrival of session stream data.Similarly,we have also added the Transformer part to overcome the limitations of the model in learning complex representations.As for the experiment,this paper adopts MRR@20 and P@20 Two evaluation indicators are tested on multiple datasets.The results show that the method proposed in this paper has significantly improved the evaluation indicators and the performance of the session recommendation system,whether in session datasets or stream datasets.The experimental results verify the effectiveness of the model and can generate more accurate recommendations for users.