| Nowadays,complex networks are ubiquitous,which range from power networks to public traffic networks,from communication networks to air transportation networks.As a consequence,the research of complex networks has been capturing increasing attention of researchers from diverse disciplines.Among these researches,link prediction,which aims to uncover missing links in static networks and foretell future connections in dynamic networks,has become an important branch due to its wide range of applications in both theory and reality.Many real world networks are dynamic ones whose structure keeps changing over time.Link prediction in dynamic networks is more challenging and complex than in static ones owing to the dynamic nature.However,effectively using the information carried by dynamic networks can make notable enhancements to the prediction accuracy.In this regards,this thesis proposes two link prediction methods for dynamic networks.(1)Link prediction in dynamic networks based on the fusion of features.Due to the evolution nature of dynamic networks,these networks have different structural characteristics at different time.This thesis proposes a dynamic network link prediction method based on TOPSIS and DS evidence theory.This method considers a variety of structure-based similarity indexes and the connection states of a node pair at different times as the features of the node pair,and uses TOPSIS to fuse the features in one snapshot to obtain the prediction result from the snapshot for the node pair.Then,the method takes advantage of the DS evidence theory to integrate the prediction results of different snapshots to obtain the final similarity score.In the process of fusion,the contribution of different snapshots is comprehensively considered.Experimental results on 8 real networks show that the proposed method achieves better prediction accuracy and is more robust.(2)Link prediction in dynamic networks based on topological and historical features.In order to further improve the accuracy of link prediction,this thesis proposes a new supervised method,named THILP.This method treats link prediction as a regression problem.In this regard,both topological and historical features are extracted from multiple snapshots to represent node pairs,and then the Random Forest Regressor algorithm is adopted to train a prediction model.Extensive experiments are executed on nine benchmark networks to investigate the effectiveness of the THILP method.The results indicate that THILP behaves remarkably better than baseline methods. |
