Research On Cross-Domain Sequential Recommendation Algorithm Based On Graph Nearal Network

The information society provides people with rich information,but at the same time,it brings the problem of information overload,which is not conducive to people to screen out useful information.The solution of this problem depends on the recommendation system.With the indepth study of deep learning,the application of deep learning model in recommendation system has also made new progress,especially the rise of graph neural network in recent years,which provides a new research perspective for sequence recommendation.On the one hand,the current recommendation system often lacks corresponding data when facing the emergence of new users or new items.In order to alleviate this problem,cross domain recommendation came into being.On the other hand,in real life,there is a phenomenon that multiple users share one account,which leads to the behavior of multiple users being recorded in a mixed account,which mixes the interests of multiple users,and brings new challenges to the recommendation system.In this thesis,by considering the behavior sequences of users in different domains,the existing graph neural network model is improved to recommend cross domain sequences.Specifically,this thesis studies a new cross domain sequence recommendation task in the shared account scenario.The task mainly faces the following two challenges:(1)the behaviors of users in shared accounts are mixed,so it is difficult to distinguish them,which makes it difficult to mine the interests of different users.(2)It is difficult to extract the relevant information of users in different fields to help the target domain,that is,extract useful information from the source domain and transmit it to the target domain to enhance the recommendation results of the target domain.The main research contents of this thesis are as follows:This thesis proposes a domain aware attention graph convolution network model.In this model,we first connect users and items in different domains by constructing a cross domain sequence diagram,and then design two new attention mechanisms to select relevant information differently in the process of information transmission,and update the representation of the current node(user / item)through different types of connected nodes,so as to realize the transmission and modeling of cross domain information on the basis of considering structural information.In addition,in the transmission process,this thesis also considers the characteristics of shared accounts,assuming that there are h potential users under each account,and uses the message transmission strategy in the domain awareness graph convolution network to transmit and aggregate messages according to users,so as to further realize the modeling of account interest diversity.Based on the above algorithm,a time interval enhanced domain perception graph convolution network model is further proposed.In this model,firstly,the construction method of cross domain sequence diagram in the above domain perception graph neural network model is used for reference to connect different types of nodes in one graph.Then,the representation of each user and item in the graph is learned through the message passing mechanism in the graph.In item representation learning,the model follows the sequence aware attention mechanism in the domain aware attention graph convolution network model to capture the different importance of different neighbor users / items.However,different from the former,in the process of item learning,we further integrate the time interval into the information transmission process,and optimize the item learning process by considering the time interval information.In addition,in order to further enhance the representation learning of sequences,we designed an account aware self-attention module to learn the interactive characteristics of items.In this thesis,the two domain-aware attention graph convolution network algorithms are proposed in the datasets HVIDEO,HAMAZON,H＿VIDEO and H＿AMAZON was verified respectively.The experimental results show that compared with the existing related methods,the two methods proposed in this thesis can achieve better experimental results on MRR and recall indexes,which proves the effectiveness of the graph-based solution proposed in this thesis.