Session-Based Graph Neural Network Recommendation Algorithm

With the wide application of deep learning on recommender systems,it is clear that the research on session-based graph neural network recommendation algorithms has become a new paradigm for recommender systems.Session-based recommendation systems aim at exploring users’ main preferences and predicting their next clicks through anonymous session data when historical user behavior data is not available and user information is incomplete.Although the existing models have achieved good prediction results,there are still some shortcomings.First,when fusing additional information such as current session order information and neighbor characteristics,existing models ignore the importance of the main purpose of the current session when fusing additional information of the session,which makes the model more noisy in enriching the current session information and reduces the recommendation performance.Secondly,most of the existing models can only perceive the item conversion information within a session,and cannot use a large amount of historical relationship information and information of neighboring items in other sessions that are common with the current session,which makes the model unable to fully explore the user preferences of the current session and has certain recommendation limitations.In this paper,we conduct an in-depth study on the session-based graph neural network recommendation algorithm and complete experiments on a real-world dataset widely used in recommendation models to demonstrate the effectiveness and rationality of the proposed method.The main work of this paper is as follows:1.We investigate the scientific literature on Markov chain-based session recommendation,collaborative filtering-based session recommendation,and deep learning-based session recommendation,and analyze the existing classical graph neural network session recommendation models in depth,and present the advanced features and shortcomings of the existing models.2.A root mean square value session information fusion method is proposed to accurately represent the main purpose of the current session of the anonymous user.Among them,the root mean square value information can more accurately represent the average baseline of all items in the session,reflecting the average value of all key nodes in the session and the difference between each item and the average baseline.At the same time,when aggregating session order information and neighbor characteristics,considering the average value of session information can mitigate the noise impact of other information on session prediction,and can also effectively solve the problem of losing order information and remote dependencies between items after using graph construction for user historical behavior sequences,as well as the problem of insufficient transformation information of current session items.3.A novel approach to fuse the features of neighboring items in the global session and the current session representation is proposed to solve the problem that the existing real-world data sessions are generally too short in length and the implied inter-item transformation information is insufficient.At the same time,the global neighbor features of each item in the current session are mined to make full use of other session information to enhance the accuracy and effectiveness of the current session item embedding and improve the recommendation effect of the model.4.Extensive experiments have been conducted on real-world benchmark datasets with a large number of ablation experiments to demonstrate the effectiveness of the two methods proposed in this paper.Practical insights are provided for the session-based recommendation problem.