Research Of Graph Neural Network Model And Its Application On Sequence Recommendation

In the era of the information explosion,recommendation system can mine and analyze the massive amount of web information to provide personalized services that meet users' needs.And there may be a large amount of non-euclidean data in the huge amount of information,such as,graph data.Graph neural networks are graph-based representation learning methods,which can aggregate information about neighbouring nodes,generate high-quality node feature representations,make predictions and classifications,and provide accurate recommendations for users.Mainstream recommendation systems make recommendation decisions based on the users' complete personal information and historical behaviour,whereas in some scenarios,users are anonymous and their interests are dynamic and immediate.Therefore,it is necessary to consider making recommendation decisions based on the interaction sequences of anonymous users,i.e.sequential recommendation.In this thesis,we investigate graph neural network algorithms applicable to sequential recommendation in order to better capture the complex transformation relationships between items,model them accurately,and apply them to recommender systems.The following work is accomplished in this thesis.1.A new graph neural network model based on lossless sequential graphs is proposed.The model includes:(1)a new lossless session graph coding,which solves the problem of sequence order information loss in the process of session transformation into graph modelling,by which the information of edges in the session graph can be improved.(2)A top-n prediction method based on a multi-headed attention mechanism,which gives precise learning parameters for top-n prediction.2.A graph neural network model based on global preference augmentation is proposed.The model includes:(1)a modeling method based on global graph and session graph,which learns users' global preferences and short-term intentions from global graph and session graph respectively,making users' preferences more accurate.(2)A top-n prediction method based on multi-headed attention mechanism,which makes top-n prediction results more accurate.In this thesis,the effectiveness of the above models are verified on four public datasets.Compared with the existing sequential recommendation baseline models,the P@20 of the two models has been improved by 13.88% and 21.67% respectively,and the MRR@20 of the two models has been improved by 20.88% and 14.77%respectively.Finally,in this thesis,the online music recommendation system is designed and implemented with the above graph neural network models as the core algorithm.The system mainly includes the functions of data processing,recommendation service and recommendation result display.