Research On Link Prediction Method Based On PU Learning

With the rapid development of network information technology represented by the Internet,people have acquired various data of complex networks.Link prediction,as an important issue in data analysis and mining of complex networks,has been widely concerned by researchers in various fields.Link prediction can analyze changes of the network structure by using theoretical basis of data mining,and research the problem that whether the target link will form or not between two nodes in the future.The traditional methods of link prediction are mainly designed for homogeneous networks with the single type of nodes and links.However,the real world networks are mostly heterogeneous with multiple types of nodes and links,which lead to more complicated relations between node pairs.Considering that the relation between a node pair can be represented not only by a direct link,but also by a path that mixes various nodes and links in heterogeneous networks.Thus,the link prediction problem can be extended to the relation prediction problem.Link prediction is mostly regarded as the binary classification problem under the supervised learning framework.The node pair has the target link is labeled as the positive instance in the network.Otherwise,the negative instance.In fact,these labeled negative instances are not credible(target links between some node pairs may form in the future)and maybe reduce the performance of the prediction model.This paper attempts to study the link prediction problem from the perspective of PU learning in complex information networks.The node pairs without the target link/relation are considered as unlabeled instances rather than negative instances.Due to most challenge problem that the serious imbalance of data number between the positive set P and unlabeled set U,we study how to extract the reliable negative set RN from set U.Thus,we propose a K-means and voting mechanism based technique SemiPUclus.Then we implement a link prediction framework PULP of homogeneous information network based on PU Learning and a relation prediction framework PURP of heterogeneous information network based on PU Learning.Finally,experimental results show that PULP and PURP achieve better performance than comparative methods in DBLP and Twitter networks.