Design And Management Of Link Prediction Model For Social Friend Recommendation

With the rapid development of internet technology,online social networks have become an important platform for people to communicate.The recommendation of social friends can expand the user’s social network,help users find friends with similar interests,and obtain a better social experience.In social networks,users not only have their own basic attribute information and interest content information,but also have social friend network topology information and heterogeneous social behavior activity information such as browsing,liking and forwarding.Based on the above complex information characteristics of social users,this thesis studies the link prediction algorithm of social networks based on homogeneous graph neural network and the link prediction algorithm of social networks based on heterogeneous graph neural network to effectively mine user content attribute characteristics and network topology characteristics.Based on the above research,a link prediction model management system for social friend recommendation is designed and implemented.The main work done in this thesis is as follows:(1)For friend recommendation scenarios with only user friend network topology,this thesis designed a link prediction model based on overlapping neighborhood and graph neural network model design space.This model first generates the optimal graph neural network model for the target dataset to learn the node feature representation through the graph neural network model space search combination.Then the topological structure features between the predicted node pairs are obtained by learning the prediction edge overlapping neighborhood features.The prediction edge scores of the two features learned above are combined with the target user-level attention mechanism to achieve end-to-end link prediction for friend recommendation.In many public social network datasets,such as BlogCataLog,Twitter,Slashdot,etc.,the link prediction algorithm proposed in this thesis outperforms the current popular link prediction algorithms in friend recommendation tasks,with a maximum accuracy improvement of 4%.(2)For friend recommendation scenarios with multiple attribute features and various social behavior network topologies,this thesis proposes a multi-dimensional user feature fusion link prediction model based on heterogeneous graph neural networks.The model utilizes a heterogeneous graph neural network to learn various social network topology features and heterogeneous data information of users.This model also concludes a multi-dimensional user feature attention algorithm that weights the different dimensions of user features,fusing the learned multidimensional user representations into a final user vector representation for social friend recommendation.In Tencent Weibo and LDBC-SNB datasets,the model proposed in this thesis achieves the best recommendation effect.(3)In view of the variety of social network topologies and link prediction models in actual scenarios,and the lack of standardized management of link prediction models,this thesis designs and implements a link prediction model management system for social friend recommendation.The system is based on the research of the link prediction model above,and mainly implements the functions of graph dataset management and analysis,model design,management and training,model log information management and analysis,model automatic parameter tuning,and friend recommendation.This system can help users analyze the structural characteristics of social networks,quickly realize and manage the optimal link prediction model,and users can use the optimal recommendation model to quickly recommend social friends.