Algorithm Research On Session-Based Recommendation System

In recent years,with the rise of streaming media,a large amount of user access data presents fragmented characteristics,where users interact with applications at a high frequency within a short period of time.Traditional collaborative recommendation systems,which rely on user-item matrices,cannot effectively model different user access sequences and generate differentiated recommendations towards different sequences.As a result,session-based recommendation systems have been proposed,in which the sequence of items accessed by users within a short period of time is defined as a session.This dissertation mainly investigates the problems of session-based recommendation systems in different application scenarios.First,anonymous session-based recommendation systems.They are defined as predicting the next item a user might click based solely on the item access sequence in the current session,without knowing the user’s identity information.The state-of-the-art algorithms model sessions as session graphs and global session graphs,capturing the transition information within them.However,this dissertation finds that existing algorithms cannot effectively extract session intention information from the global session graph that motivates the generation of sessions.To address this issue,this dissertation proposes the concept:the global session neighbor set,which can effectively express session intention information.To improve the model’s ability to capture session intention information,this dissertation introduces a supervised contrastive learning method.Then,this dissertation proposes to combine this supervised contrastive learning task with the recommendation task to construct a multi-task learning framework,RGANSBR(Related GlobA1 Neighbor enhanced Session-Based Recommendation system).Extensive experiments on three real-world datasets validate the superiority of our proposed RGAN-SBR model compared to existing models.Second,personalized session-based recommendation systems.They remove the assumption of anonymous source users in sessions and can be applied to a wider range of scenarios.They are defined as predicting the next item a user might click based on the user’s historical session information and the item access sequence in the user’s current session.There exists a natural "user-session-item" hierarchical set structure in this problem,which existing methods fail to utilize effectively,resulting in significant information loss.To address this problem,this dissertation proposes a hierarchical hypergraph neural network to model this structure holistically.Furthermore,this dissertation introduces a directed graph aggregation model to learn structural information in directed global graphs,which complements the topology relationships of items in one session that the hierarchical hypergraph neural network cannot model.This dissertation proposes the H3GNN(Hybrid Hierarchical Hypergraph Neural Network)model to combine the item representation vectors generated by the two models.Comparative experiments on three real-world datasets validate the superiority of our proposed H3GNN model compared to existing models.In summary,anonymous session-based recommendation systems first introduced modeling methods of session-type data and can be applied to streaming applications without a user base and recommendation scenarios with anonymous website visits.Personalized session-based recommendation systems,on the other hand,were inspired by the modeling method of session-type data of anonymous session-based recommendation systems.They were proposed to get rid of the limitations of the anonymity premise in anonymous session-based recommendation systems,and thus,they can be widely applied to streaming applications with a certain user base.These two problems belong to session-based recommendation systems for different types of streaming applications in different development stages,and studying them can effectively provide a profound understanding of the panorama of streaming applications.