Time And Structure Enhanced Multi-Channel Graph Neural Network For Session-Based Recommendation

With the vigorous development of the Internet,when users use e-commerce platforms or video streaming platforms on the daily life,they will inevitably face the problem of information overload.How to recommend products or videos that meet the user’s preferences from a large amount of data is the main problem to be solved by the recommendation systems.Recommendation systems are currently used in a variety of scenarios.Although the recommendation system has become an effective tool for resolving information overload,while in session-based recommender systems(SBRSs)scenarios,long-term historical behavior information and some user-side information cannot be obtained because it is targeted at unlogged-in users or anonymous users(only short-term user behavior records are included).This makes the general recommendation system unable to satisfy users’ needs by relying only on users’ short-term behavior information.As a result,the research on sessionbased recommendation has received more and more attention from academia and industry.Session-based recommendation aims to model user preferences based on the user’s shortterm behavior records to achieve the purpose of personalized recommendation.This thesis mainly studies the session-based recommendation task from the perspective of time and the sparseness of session data and studies the session-based recommendation model of multi-channel graph neural network with time and structure enhancement.This study builds two session-based recommendation models and conducts several experiments to verify the validity of the models.The main contents of this thesis are as follows:Time Enhanced Graph Neural Network for Session-based Recommendation.For session-based recommendation tasks,previous studies have focused on modeling complex transformations between non-adjacent items by capturing sequential transformations between consecutive items by using a recursive neural network(RNN)or a graph-based neural network(GNN).Although these efforts have achieved encouraging results in resolving session-based recommendation issues,while few people are committed to exploring rich information related to the changes of user interest in transitional relationships,which is the research gap that this thesis attempts to fill in.In this work,a new time-based model called Temporal Enhanced Graph Neural Network for Session-based Recommendation(TE-GNN)is proposed,which captures complex user interest transfer patterns in a session.TE-GNN constructs a Temporal Enhanced Session Graph(TES Graph)which adaptively generates conversion relationships between items based on the degree of user interest drift.In addition,TE-GNN has designed a novel Temporal Graph Convolutional Network(T-GCN)to learn items embedding.Finally,TE-GNN models the representation of items with common user interests by introducing a Temporal Interest Attention Network(TIAN).Experiments on four benchmark datasets show that the TE-GNN presented in this thesis is significantly better than the most advanced baseline method.Dynamic Global Structure Multi-channel Enhanced Graph Neural Network for Session-based Recommendation.Most current session-based recommendation methods only use the current session sequence to model user preferences,ignoring the rich information from a global perspective.Meanwhile,previous advanced works often used GNNs to capture conversion relationships between items,but the graphs used in GNNs are constructed in a static mode,which may introduce noise to the graph structure if user preferences changed.In this work,a Dynamic Global Structure Enhanced Multi-channel Graph Neural Network(DGS-MGNN)is proposed for session-based recommendation.DGS-MGNN presents a new GNN model called Multi-channel Graph Neural Network(MC-GNN).MC-GNN can dynamically generate local,global,and consensus graphs and represent more information based on the corresponding graph learning items.At the same time,to reduce the noise information in the session,DGS-MGNN uses a graph structure to assist the attention mechanism to filter the noise information in the session so as to generate accurate intent representation for the user.Finally,a more accurate prediction probability distribution is generated by combining repeat and explore modules.Experiments on three widely used datasets show that DGS-MGNN is always superior to the most advanced baseline model.