Research On Recommendation Based On Graph Neural Network And Self-Attention Network

Session based recommendation aims to predict the user’s next click on an item based on the user’s current session and historical behavior sequence.Most existing session based recommendation systems only predict user behavior by establishing short-term interest preferences for the current session,often ignoring the long-term interest preferences of the user’s session,thus underestimating the information contained in the global sequence of the session.In addition,the location information transmitted by users when clicking on items represents different interest changing processes for users,but most recommendation systems do not consider the relative location relationship between conversation sequences.To address the above issues,a session recommendation model based on the fusion of graph neural networks and improved self attention networks(Self-Attention Graph Neural Network,GNN-SA)is proposed to jointly consider changes in users’ short-term and long-term interest preferences.GNNSA introduces graph neural networks and traditional attention mechanisms to extract context information in local conversations to capture users’ short-term interest preferences;The self attention network is improved as a preference learning module to extract global context information to capture users’ long-term interest preferences.By linearly fusing the two,the next recommendation prediction is provided to the user.Secondly,on the basis of the GNN-SA model,a self-learning location embedding module,a feedforward neural network,and a residual network are added to the nodes of the user’s historical session to form GNN-SAP(Self Attention Graph Neural Network with Position Embedding,GNNSAP).By adding location information to enhance the representation ability of the model,It can also extract nonlinear information about users’ long-term preferences and alleviate problems such as gradient disappearance and gradient explosion during their training process,improving the performance and accuracy of the recommendation system.By comparing GNN-SAP with benchmark models commonly used in the recommendation field on different datasets,it is verified that it has significant optimization improvements over existing session recommendation methods on commonly used sparse and dense datasets and different evaluation indicators.Finally,based on the GNN-SAP model,a personalized music recommendation system based on users is designed and developed.The system design is concise,easy to operate,and the user interface fits the user’s usage habits.It makes full use of user history playback records,collection records,and download records to capture their long-term and short-term song preferences,and customizes and recommends songs for users,which can greatly improve the user experience.