Research On Session-Based Recommendation Model Based On Graph Neural Network

In recent years,recommender systems have received extensive attention,and sessionbased recommendation is one of the research hot spots.There is a lot of research in the field of session-based recommendation and various types of session-based recommendation models have been proposed.Among them,graph neural network-based session-based recommendation models have recently achieved the best recommendation performance.However,these models still have some shortcomings.Firstly,most of the existing sessionbased recommendation models do not fully mine the two types of information between items in the session,which limits the representation of items and sessions.Secondly,modeling the two types of information between items can be seen as two subtasks of session-based recommendation,and using the multi-task learning method can explore the relationship between tasks and improve the performance of the model.In this thesis,we conducted in-depth research and proposed two new session-based recommendation models.We also verified the effectiveness of the proposed models on widely used real-world datasets.The main work of this thesis includes:1.Investigating the literature related to session-based recommendation and graph neural networks,and conducting analysis of existing session-based recommendation models to point out possible improvement methods.2.This thesis proposes a session-based recommendation model named RelevanceAware Graph Neural Network(RA-GNN).There are two types of information between items in sessions,i.e.,the information of item-to-item relevance and transitions,but existing session-based recommendation models generally only use the information of item-toitem transitions.Thus,this thesis proposes RA-GNN.First,a method for constructing a relevance graph is proposed,and the graph convolutional network is improved to design a relevance embedding module.The relevance embedding module model the relevance information by importing the relevance graph.Secondly,a dual-information utilization mechanism is proposed,which constructs different graphs and embedding modules for the two types of information according to the input sessions,which further mines the information between items.RA-GNN is compared with existing session-based recommendation models on real-world datasets,and the results show that RA-GNN can improve the performance of session-based recommendation.3.This thesis proposes a session-based recommendation model named MML-SR.Existing studies have shown that multi-task learning can effectively improve the performance of machine learning tasks.Based on the idea of multi-task learning,this thesis proposes a session recommendation model,MML-SR,which builds upon RA-GNN.MML-SR divides a session-based recommendation task into a relevance recommendation task and a transition recommendation task,and then designs expert network layers and gate neural network layers to capture the interaction of different information in the task and explore the relationship between tasks.the MML-SR model also dynamically adjusts the weights of different tasks to distinguish the impact of different types of information on sessionbased recommendation.MML-SR is compared with existing session-based recommendation models on real-world datasets,and the results show that MML-SR can improve the performance of session-based recommendation.