Research On Link Prediction Model Based On Structural Attributes Of Complex Networks

Link prediction is one of the major research directions in information technology field of 21st century,it is an interdisciplinary area combining data mining,information science,topology,artificial intelligence and complex network.It can not only be applied to conventional recommendation systems,interactive search,but also to new applications such as link maintenance and behavior pattern analysis in 6G networks.All above applications require high accuracy of link prediction.Nowadays,one of its research directions is to obtain the possibility of connection between the nodes by quantifying their influence.One of the key problem is how to fully and efficiently utilize the structural information of the network.Therefore,this thesis researches on the related models based on the complex network structure attributes as influences.The main contributions include:(1)Aiming at the problem of low accuracy of link prediction models caused by insufficient utilization of endpoints information,this thesis first explores the feasibility of using node degree and H-index as node influence indicators on Local Path(LP).Then,the Degree Influence of Local Path(DILP)model,the H-index Influence of Local Path(HILP)model and the Degree and H-index Influence of Local Path(DHILP)model are proposed respectively.The simulation results on 12 real datasets shows that compared to the LP model,the DILP model,HILP model,and DHILP model improved the performance of link prediction,confirming that applying both node degree and H-index as influence can improve the prediction accuracy under local paths.What’s more,the model applying above two attributes as influence is better than the model with single one.(2)In different types of networks,the contributions of multiple attributes of nodes to influence varies.If the influence of neighbor nodes only considers a simple mix of structural attributes,this problem limits the link prediction accuracy and the scalability of endpoint influence.To this end,this thesis proposes a Weighted Influence of Neighbors(WIN)model based on neighbor degree and H-index.This model considers the random walk process,and appropriate weights are obtained to make the prediction accuracy achieve the best performance by setting the weights of the degree and H-index on different networks.The simulation results on 12 real datasets verify the effectiveness of the proposed model.(3)Through in-depth analysis of the intrinsic relationships between node attributes in existing Hybrid Influence of Neighbors(HIN)models,it is found that the coreness contains more information than the H-index.Therefore,a hybrid influence of neighbors(DCHIN)model based on neighbor degree and coreness is proposed.The simulation results on 12 real datasets shows that compared with HIN model,DCHIN model improves the performance without increasing time complexity.