Research On Session-based Recommendation Model Based On Long And Short-term Interest And Neighborhood Information

With the development of big data and other information technologies in recent years,the exponential growth of Internet resources has led to the disaster of information overload.How to let users get the information which meets their personal needs from massive resources accurately has become a huge challenge for the Internet industry.As a decision support technology in the field of artificial intelligence,recommendation system solves this problem well.Most recommendation systems can make personalized recommendations for users by inferring their interests and preferences based on personal information and historical behavior.At present,the research on session-based recommendation is of great significance in recommendation technology,because the system usually cannot obtain complete personal information and long-term historical behavior data of users in real application scenarios.The session-based recommendation method does not need to access personal information,but only completes the recommendation according to the interaction sequence between users and items in the session.With the development of deep learning technology,session-based recommendation has achieved remarkable results in practical applications,however,the existing session-based recommendation model still has the following problems: in the process of modeling users’ interests,most models do not consider that the importance of long and shortterm interest in generating the users’ overall interest is different in sessions of varying lengths;most algorithms only use the limited information in the current session to build models,which will cause insufficient use of neighborhood information in the global session relationship.In order to provide users accurate personalized recommendations more effectively,this paper proposes corresponding improvement strategies based on the existing research work to solve the above problems,the main work is as follows:(1)This paper designs a session-based recommendation model based on adaptive fusion of long and short-term interests(SR-AFLSI).Firstly,the model uses a gated graph neural network integrating self-attention mechanism to update the node information in the session graph and obtain the item embedding representation.Secondly,the session length is normalized to generate the long and short-term interest fusion weights,and the long and short-term interests are aggregated by weighted fusion to obtain the session embedding representation that represents the overall interest of the user.Finally,the model completes the recommendation according to the session representation.(2)This paper designs a session-based recommendation model with graph neural network based on neighborhood information enhancement(SR-NIE).It not only models the current session,but also introduces the neighborhood information in other sessions as a supplement to achieve the purpose of information amplification.Firstly,the model constructs the session graph and the position-aware global graph and takes into account the relative position information between items in the position-aware global graph.Secondly,the model obtains the item embedding representation by learning two kinds of graph structures.Then,the session embedding representation which contains neighborhood information is obtained by the attention mechanism which aggregates item reverse position embedding and user long and short-term interest representation.Finally,the model completes the recommendation.(3)In this paper,we use real public data sets to test the designed two models.Through comparison experiments with the baseline model and self-ablation experiments,the effectiveness of the two models and the rationality of each module set in the model were verified.