Research And Implementation Of Session-based Recommendation Based On Graph Neural Network

With the popularization of mobile smart terminals and the rapid development of the Internet industry,people can easily obtain information and make decisions on the Internet.However,the explosive growth of users and items has brought huge challenges to information matching.In order to alleviate the serious problem of information overload,service providers need to design recommendation algorithms to mine users' interests and preference,so as to filter information that user may be of interest.The session-based recommendation method proposed in this paper is to predict the user's next click behavior based on the user's item click sequence over a period of time,so as to assist user to make decision.The complexity of user composition makes it difficult to predict user behavior.In order to better analyze user click behavior,this paper designs a session-based recommendation system based on graph neural networks,and uses data enhancement to improve the original system's ability to handle different situations.Specifically,1)To solve the problem of ignoring the global transfer relationship of items in the traditional method,this paper proposes a session-based recommendation method based on graph neural network,which aggregates the context information and item transfer relationships into one session graph.Then user embedding is extracted through the gated graph neural network and the attention mechanism,so that a better recommendation list can be generated considering the context information and the global item transfer relationship.2)To deal with the uncertainty of the behavior of users and businesses,this paper proposes a embedding enhancement method,which augments the original user conversation graph through node dropping,edge perturbation and attribute masking.This method improves the robustness of the original model,and allows it to handle various abnormal data.3)To solve the problem of insufficient exposure of newly released items,this article innovatively proposes a cold-start training method for item embedding based on graph embedding,using item attributes and random walk strategies to optimize the initialization of new item embedding.In terms of experiments,this article uses Recall and MRR as evaluation indicators.The results on the data sets RecSys2015 and LastFM show that the method proposed in this article is significantly better than the widely used methods in session-based recommendation scenarios,and can provide users a more suitable recommendation list.This article also developed a recommendation prototype system,and verified the feasibility of the application of the session-based recommendation method based on graph neural network.