Link Perspective Based Network Embedding for Link Prediction

Abstract

Link prediction is an important task attractingmuch attention in the field of complex network, and it can be applied to many real-world scenarios such as recommendation engines, protein-protein interactions prediction, and stock index prediction. Recently, network embedding, which learns low-dimension latent representations of vertexes, has attracted considerable research efforts. It provides a new feasible solution to boost the accuracy of link prediction by preserving the rich structure information of the network. Most existing methods for network embedding depend exclusively on the perspective of nodes, but seldom focus on the perspective of links. In practice, different perspectives of network can bring us different information. As our empirical analysis shows, random walking from node perspective and link perspective of network could bring us different node sampling results, which may mean different partitions of communities. Thus, we designed a new network embedding approach for link prediction which incorporates the perspectives of both nodes and links. The aim of our method is to supplement the structure information from the perspective of links by using line graph, which improve the validity of nodes representations for link prediction. We compare our method with other popular approaches on four real-world networks and the experiment results demonstrate that our method improves performance in three link tasks.