Research On Personalized Recommendation Based On Deep Learning Link Prediction

Link prediction aims to reveal the underlying relationships behind the network,and it has important research value as a study in the field of data mining.Personalized recommendations,as a very important application of link prediction,have also received a lot of attention.Link prediction to eventually personalize recommendations is a critical way to solve such problems.Link prediction recommendation method as a commonly used recommendation method,that is,on the basis of the known part of the network node information and structural information,how likely or how likely there will be a link between the two nodes in the network that have not yet produced a connection edge,and to predict and implement the recommendation.Graph attention network can consider the correlation and correlation of nodes and their neighbors at the same time,which can effectively improve the accurate extraction efficiency of node feature vectors.In order to improve the accuracy of personalized recommendation,this thesis proposes a link prediction personalized recommendation method based on multi-graph attention network.This method uses multiple graphs for embedding,updating initialization feature vectors,getting the feature vector after multi graph fusion,and performing feature learning through graph attention network,and the output features are used to predict scores,so as to achieve link prediction and finally personalized recommendation.The embedding of multiple graphs can express the higher-order information of the model and integrate adjacent information.The two-layered attention mechanism set up in the graph attention network obtains the potential components of the user’s(project)characteristics,and aggregates these components to automatically learn the importance of different components.The method used in this thesis uses RMSE and MAE evaluation indicators,and through comparison with some mainstream algorithms,the experimental results of three real data sets show that the proposed method can achieve better personalized recommendation effect.The main research contents of this thesis are as follows:(1)Construction of multiple graphsIn the process of embedded learning,multiple graphs are included for embedded learning at the same time,because the existing methods rarely consider the influence of the construction of multiple graphs on improving personalized recommendation,and do not make full use of user-item historical interaction data information and user-user and item-item relationships for personalized recommendation.The multi-graph can comprehensively consider high-level information and adjacent information,the user-item bipartite graph can express the high-level information of the model,and the user-user and item-item similarity graphs can integrate adjacent information.The combination of high-level information and adjacent information is fully considered,so that the learned user and project characteristics are more comprehensive,and the prediction effect is further improved,so that the personalized recommendation model is more accurate.(2)Graph Attention NetworkThe graph attention network is formed by introducing an attention mechanism based on graph convolution.Graph convolution cannot treat each user differently,but the attention mechanism can assign different weights to nodes,and different weights represent different importance of nodes.In this paper,the graph attention network is used to conduct research projects,which can not only consider the structural information between nodes and nodes at the same time,but also automatically learn the importance of different nodes.(3)Link predictionThe key to link prediction is to predict the probability of its occurrence,that is,the probability of its occurrence.In this paper,a link prediction method based on similarity is used to make link prediction more accurate.The value of RMSE and MAE is used to determine the effect of the link prediction.Then,the results of the link prediction are used for personalized recommendations,and the more accurate the results of the link prediction,the more accurate the personalized recommendations,and the more satisfied people are with the recommendation results.(4)Personalized recommendationAfter learning the final features of the user and the item,the predicted score is calculated through the multi-layer perceptron,and the predicted score is compared with the real score.The closer the score is,the better the link prediction effect is,and the personalized Recommendations are more accurate.Using RMSE and MAE as evaluation indicators,and comparing with some mainstream algorithms,it is concluded that the proposed method can achieve better results in terms of the accuracy of personalized recommendations.