Research On Social Network Link Prediction Algorithm Based On Tensor Factorization

With the wide application of social network data analysis technology,the demand for research on social network is increasing.However,the scale of social networks continues to expand,and the information to be exchanged increases exponentially.Therefore,how to complete and discover link information in social networks has become an important research direction of common concern in both academia and industry.Among them,link prediction is one of the important means to study social networks.Traditional link prediction algorithms focus on the discovery of historical static links,without considering the dynamic characteristics of social networks.Therefore,in order to solve the limitations of traditional link prediction algorithms,this thesis proposes an algorithm form to improve the accuracy of social network link prediction and further expand its applicability.This thesis adopts the method of non-negative tensor factorization,innovatively incorporating the principles of linear bias and alternating direction method,and reconstructing the parameter learning rules,a Biased non-negative tensor factorization(BNTF)model is proposed.Transforming the dynamic social network link prediction problem into an incomplete tensor missing value prediction problem effectively improves the representation learning of temporal features of social networks.Experiments show that the model achieves 2.1%-10.3% reduction in prediction error compared to state-of-the-art missing link predictors.Then,in order to solve the problem that the BNTF model is not suitable for predicting future links in dynamic social networks and further expand its applicability,proposes a time series prediction model based on biased non-negative tensor factorization.By applying the Holt-Winters(HW)time series prediction algorithm to the temporal latent features extracted by the BNTF model,it innovatively predicts the future links of the entire social network with a single latent feature trend.Experiments show that the model achieves the highest AUC value compared with state-of-the-art future link predictors,further extending the applicability of the BNTF model.In summary,the model proposed in this thesis can more accurately represent the trend of latent features in social networks over time,and achieve high-precision prediction of social network links.By combining the Holt-Winters time series forecasting algorithm,it shows that other time series forecasting algorithms may also be applied to the BNTF model,which has great application potential in different social network tasks.