Research On User Trust Prediction Methods In Social Networks

The rapid development of the Internet and the increasing popularity of mobile devices have made social networks a major platform for obtaining information and interactive entertainment,and more serious security issues have followed.At present,in addition to the traditional hard security methods,the trust mechanism,as an important soft security method,has been widely used.Trust and distrust relationships among users play an important role in social network applications,for example,trust relationships can enhance social recommendation,and distrust relationships can be used for fraud detection.However,existing trust relationships in social networks are very sparse and cannot provide rich trust information.In order to apply trust relationships to different applications in social networks,it is necessary to predict the possible trust relationships among users.Trust prediction aims to predict the trust relationships between any two users,which is the key to helping users discover reliable information.This dissertation focuses on the research of trust prediction methods in social networks,especially on improving the accuracy of trust prediction and robustness to trust relationship sparsity.Three trust prediction methods are proposed.The research contents and contributions of this dissertation are shown as follows:(1)Aiming at the low accuracy of propagation-based trust prediction methods,a multi-factor deep trust prediction method(G-Deep Trust)is proposed.In this method,various features such as user attributes,network structure,and trust relationships are integrated into the trust prediction process,and user attribute features are constructed from four aspects based on social trust influencing factors.The dissemination,aggregation,and asymmetry of trust are comprehensively considered,and a hybrid structure based on the graph convolutional networks and graph attention network is designed to aggregate vector representation of neighbor nodes from two aspects:popularity trust and activity trust.Compared with other existing methods,the prediction accuracy of the method proposed in this paper is higher.At the same time,the rules of trust propagation and aggregation are effectively captured,and the generated user potential trust factors are also more accurate.(2)Aiming at the problem that the performance of existing trust prediction methods is affected by data sparsity,a trust prediction method based on deep learningenhanced matrix factorization(DM-Deep Trust)is proposed.In this method,the influence of homophily is considered,and the users’ preference similarities are used as a regular term of matrix factorization.In order to better evaluate the users’ preference similarities,a deep learning language model is used to obtain text vector representations from the user comment text,which are used as the users’ preference features.Then the feature vectors are dimensionally reduced using a deep autoencoder,and similarities between feature vectors are calculated.Finally,a homophily regularization term is designed to make it easier for similar users to establish trust relationships.Experimental results show that the proposed method achieves higher prediction accuracy and is robust to trust relationship sparsity.At the same time,it is further verified that the users’ preference similarities play an important role in alleviating the sparsity of trust relationships.(3)Aiming at the problem that the trust prediction method is affected by the authenticity of user context information and user subjectivity,a deep trust prediction method based on sentiment polarity perception(S-Deep Trust)is proposed.In this method,sentiment labels are applied to the trust prediction process for the first time.First,transfer learning is adopted to generate sentiment labels for user review data and obtain a sentiment polarity matrix.Then,in order to solve the problem of inconsistency between user ratings and reviews,and effectively lower the influence of user subjectivity on trust prediction,a rating matrix generation algorithm MTM is designed.The original rating matrix and sentiment polarity confidence matrix are weighted and aggregated to obtain a new rating matrix.Next,the Siamese network architecture is introduced into the field of trust prediction,and the trust relationships of users are judged from the extracted user preference vectors,which minimizes the representation of trusted user pairs,so that the trust relationships between users can be better predicted.The experimental analysis results prove that,compared with other existing methods,the method proposed in this paper achieves higher accuracy in predicting results and is robust to the sparsity of trust relationships.