| Recommendation systems can help people get personalized information and are widely used in web search,e-commerce and other fields.Traditional recommendation algorithms use users’ long-term interaction data to obtain users’ preferences for recommended items.However,in many e-commerce platforms,since users are anonymously logged in,there is no way to record their personal information and long-term interaction information.In this case,the traditional recommendation algorithm is no longer applicable,so a new branch of recommendation system is created:session recommendation.In many online systems,user interactions are organized into sessions,i.e.,a series of user behaviors that occur over a short period of time.Because user interactions in a session are always accompanied by randomness and only limited information is available to them,making recommendations in a session is challenging.Session recommendation aims to predict the user’s next click item according to the user’s previous interaction behavior in the session without the user’s historical information.At present,the application of graph neural network can obtain excellent recommendation effect in the field of session recommendation.The user’s ongoing session is modeled as a graph,and the transformation mode between items is simulated by graph neural network,so as to predict the user’s next click item.In this paper,the graph neural network theory technology is used to deeply study the session recommendation based on user behavior,and the shortcomings of the current model are improved to improve its recommendation effect.The research work of this paper is as follows:(1)To address the problem that the graph neural network-based session model has limited expressiveness due to the last item of the session representing all the interests of the user,this paper proposes a session recommendation based on graph neural network integrating session information and candidate items(SICI-GNN).It replaces the last item embedding vector with the candidate items embedding vector.Through the target attention mechanism,the process that users often choose only part of the information in the recommendation list in reality is simulated to enhance the model expression ability.Also,the sequence of items for a session is modeled as a session graph,which is used as a long-term preference to generate a more accurate representation for each item.(2)Further refinement of the SICI-GNN model for its shortcomings.The SICI-GNN model suffers from two problems: first,because the session graph processes each item separately but fails to capture the current session semantics from a high-level perspective,resulting in a large amount of information loss;second,because candidate items can contain a large amount of noisy information and increase the computational difficulty.For the former,the learning multi-granularity consecutive user intent unit for session-based recommendation(MSGIFSR)are used to solve it.The graph constructed by the multi-granularity intent unit captures intentions at different granularity levels to reduce information loss.For the latter,cross-session graphs are built with the help of cross-session information,and graph neural networks are used to aggregate all neighboring item features of the current session item,aiming to more accurately mine all the current intentions of the user and reduce the computational effort.Therefore,finally,this paper proposes a graph neural network session recommendation combining cross-session information and multi-granularity intention units.The model is improved on the basis of MSGIFSR model by adding cross-session information and combining the advantages of both in order to improve the recommendation effect.Extensive experiments are conducted on two public and real datasets,and the mean reciprocal rank and precision are used as evaluation metrics to demonstrate the validity and rationality of the proposed model in this paper. |
