Research On Disappearing Link Predictability In Soical Networks

Social network data contains great commercial value and massive social behavior information,which makes social network data mining a hot research areas recently.Link prediction,aims to predict the emergence and disappearance of links in network evolution,is a core problem in social network data mining.However,most of existing work focuses only on prediction of unknown links and future links,and there is little research on prediction of link disappearance.Firstly,based on a sociological analysis on the general phenomenon of link disappearance,prediction indicators for link disappearance in the general network scene concerning only the topology,weight and timing is proposed in this paper.Then,the correlation between the topological information and the predictive indicators is revealed by means of correlation analysis.At last,a new network evolution model considering the link disappearance-LWBA model,is proposed and proved.The main contributions of this paper are as follows:(1)Empirical study on the phenomenon of link disappearance:the prevalence of the phenomenon of link disappearance is confirmed on the basis of real-world social network data;and in further experiments,the influences of different topological factors and timing on the distribution of link disappearance are compared.(2)Method of predicating the link disappearance:the mechanism of link disappearance is explained from the perspective of social theory,and prediction indicators for link disappearance in the general network scene concerning only the topology,weight and timing is proposed.Homogeneity theory,three-degree influence theory,priority connection theory,social comparative theory,links contact strength theory and clustering theory are used.Result shows that similar to the classic future link prediction work,the link disappearance is predictable,while the prediction indicators varies in.effectiveness depending on the type of network.(3)Revealing the correlation between topological features and predication indicators using multivariate correlation analysis:To study the effect of different prediction indicators in different types of networks,the multivariate linear correlation analysis method is used to analyze the correlation between topological information and link disappearance predictors.Result shows that there is no topological feature can dominate the prediction on disappear link.(4)A novel network evolution model considering link disappearance is proposed:Based on the classical network evolution model BA,the LWBA model is proposed,which takes into account the disappearance behavior of nodes and links.The experimental result shows that the LWBA model conforms to the scaleless network.This work is a supplement to the existing research on link prediction.Result indicates that:(1)Link disappearance is prevalent in social network,and links which have reciprocal nature,higher interactive frequency and longer survival time within a certain range enjoy lower probability of link disappearance,while embedded factors have no significant influence.(2)On the basis of the general network topology,the social theory has different effects on the prediction of link disappearance in different networks.The link contact strength theory is outstanding in non-directional networks,while the social comparison theory perform better in directional networks.Additional information of weight and timing plays a positive role.(3)Using multivariate linear correlation method,the correlation between network topology features and prediction effect is analyzed through normalization,variable selecting,multicollinearity judgment and ridge regression,but no uniform conclusion was drawn.(4)According to the experimental analysis of the number of nodes and links,the average degree and the degree distribution of the LWBA model,the theoretical argument coincides basically with the experimental simulation,and the degree distribution fits power-law distribution.