Disappearing Link Prediction In Scientific Collaboration Networks

The scientific collaboration network contains rich network science information,and it is a significant subject in the field of network science to excavate the law of complex network evolution.Link prediction is an effective method for the study of network evolution.It is a common sense that both the formation and dissolution of links are the essential processes of link dynamics in network organization.The current studies focus on predicting the formation of links,but little attention has been paid to the disappearing link prediction problem.In this thesis,I investigate the disappearing link prediction problem in scientific collaboration networks in order to understand the mechanism of network evolution more deeply.Firstly,the concept of disappearing link prediction in the scientific collaboration networks is defined,and the possible factor affecting the disappearance of links are discussed.I analyze the relationship between multiple factors and the disappearance of links,and give the correlation results.Then,I study the disappearing link prediction problem from two perspectives,and propose two novel methods to solve it.The methods I adopt are based on structural similarity method and network representation learning method,respectively.I propose a structural similarity method,Modified Preferential Attachment(MPA),which takes account of the different influence of different neighbor links of the nodes.This method is simple and can quickly predict the disappearing links in the networks.I propose a network representation learning method named Time-aware Disappearing Link to Vector(TDL2vec),which combines network structure information and time attribute.This method can flexibly predict the disappearing links in the networks at different times in the future.In order to verify the performance of two methods in disappearing link prediction,these two methods are applied to DBLP and APS real-world scientific collaboration networks.The experimental results show that compared with many classical link prediction indices,MPA index has better prediction effect in disappearing link prediction problem.TDL2 vec algorithm has stronger superiority in disappearing link prediction problem.Compared with structural similarity link prediction methods,TDL2 vec can predict the disappearing links at different future times more accurately and achieve better prediction performance.