Research On Session-based Recommendation Model Based On Graph Representation Learning

With the wide application of digital devices and the rapid development of internet infrastructure,the amount of data generated by people in daily life is growing exponentially.However,information overload occurs when people receive more information than they can process.To address this problem,recommendation systems have been widely applied to help users efficiently obtain target information from massive data.Recommendation systems have broad application scenarios and significant commercial value,and have been widely used in e-commerce,social media,and other fields.However,due to people’s increasing attention to privacy protection and related policies,it is often difficult to obtain sufficient user behavior data and user profiles in many practical scenarios,leading to a decline in the performance of traditional recommender systems.To solve this problem,session-based recommendation systems have been proposed and attracted extensive attention because of their strong ability to capture users’ dynamic and short-term preferences.Session-based recommendations recommend the next item to the user by modeling the dependencies between items in a anonymous session.Due to graph neural networks are good at modeling complex relations,this thesis applies this technology to session-based recommendation to improve recommendation performance.To address the issue that existing graph neural network-based session-based recommendations often construct unreasonable session graphs,this thesis proposes a review-enhanced session-based recommendation method(RI-GNN).This method effectively utilizes item reviews to enhance the construction of session graph,filters out false dependencies between some adjacent items,and captures the true dependencies between non-adjacent items that existing methods ignore,thereby improving the performance of the session-based recommendation model.Based on RI-GNN,this thesis proposes a contrastive learning-based session recommendation method(RGSR)to solve the data sparsity problem in session recommendation while retaining complete session information.This method generates two different session graphs based on two views(user behavior and item reviews)to model the dependencies between items from different perspectives,achieving proper data augmentation and avoiding data sparsity caused by Dropout in self-supervised learning.To address the problem that current session recommendation lacks explicit supervision for modeling item representations based on external relations,this thesis proposes a dual-view session recommendation method based on substitutable and complementary relations(DIREC).This method not only uses explicit association rules(substitutable and complementary relations)via contrastive learning to supervise the modeling of item representation but also considers the use of association rules from both user behavior and item attribute perspectives,effectively modeling item representations.This thesis proposes three session-based recommendation methods and compares them with baseline methods on real-world datasets,using two recommendation metrics to evaluate the performance of models.The experimental results demonstrate that the proposed methods outperform existing methods in session-based recommendation,thereby verifying the rationality and effectiveness of the proposed methods in this thesis.