Research Of Session-based Recommendation With Graph Neural Network

Personalized recommendation systems have become ubiquitous in a variety of domains,including advertising,social media,and e-commerce.These systems leverage users’ historical interaction data to predict their preferences for specific products or services.Predicting user behavior within session is a complex task in network behavior modeling,as it is difficult to determine with certainty how a user will behave and the available information is often limited.The problem of personalized recommendation has gained significant attention in both academia and industry in recent years.The reason for this is that such systems provide users with valuable personalized information,resulting in a significant improvement in user experience and increased business revenue.As a result,the classic problem driven by the real world has become a popular area of research.Graph neural networks are increasingly being explored as a new avenue of research in recommendation systems.However,existing graph-based recommendation methods only consider simple binary relationships and cannot model complex multi-relationship between objects.In addition,different orderings of items may create the same topological relationship,which limits the ability to obtain accurate session embeddings.At the same time,existing research focuses only on the transition relationship of item sequences,ignoring other types of user behavior,such as selecting items under specific category information.Therefore,this thesis focuses on session-based recommendation research and aims to address the above limitations:(1)To address the problem that the traditional graph neural network structure only considers pairwise item transition relationships and cannot model high-order data correlations,this thesis proposes a multi-session aware hypergraph neural network,which uses a hypergraph neural network to capture the complex multi-relationships between items and solves the limitations of traditional graph neural network structures.At the same time,combining the co-occurrence graph with the local session graph considers both intra-session and inter-session information,thereby mining the similarity between users and items and improving the accuracy of session representation.Finally,experimental evaluations on Diginetica,Nowplaying,and Tmall datasets show that the P@20 values are 53.36%,23.88%,and 34.63%,respectively,demonstrating the advanced and practicality of this method in solving session-based recommendation problems.(2)To address the common real-world scenario of providing specific category information and recommending items in that category,which is not focused on by traditional session-based recommendation algorithms,this thesis proposes a category context information aware hypergraph neural network,which uses the user-specified target category and the item’s own category attributes to push potential interest items to users in the target category,making the final results more consistent with the user’s expectations.Finally,in the specified category domain environment,this method achieves P@20 values of 44.12%,21.51%,and 51.81% on the Diginetica,Yoochoose,and Jdata datasets,respectively,demonstrating significant advantages in recommendation accuracy in the specified category domain.(3)The thesis has designed an application system using session-based recommendation algorithms based on the primary research content discussed earlier.After conducting experimental tests with the algorithm,the results showed that the system was able to alleviate the problem of information overload.