| With the rapid progress of information technology,a large number of complex systems have been springing up,with the result that the research of network science based on complex network theory has been accelerated as well.As one of the important problems in the field of network science,link prediction aims to find the missing and future links between two unconnected nodes in complex networks,which has important theoretical and practical significance.On the one hand,link prediction helps us to reveal the internal structure features of networks,and develops our understanding of the network evolution mechanism.On the other hand,it has widespread application value in various real systems and consequently helps the economic development and improves the level of social governance.Focusing on complex networks,this thesis tries to investigate the influence of topological structures on the performance of link prediction from different perspectives.Firstly,the problem of path heterogeneity in unweighted networks is respectively explored from the two different views,i.e.path weight and the interactions among paths.Then,based on the intrinsic relationship between the link degree and the prediction performance in unweighted networks,we further analyze the influence of topological weight and natural weight of links on prediction performance in weighted networks.Finally,the influence of interlayer structures of multiplex networks on the prediction performance is investigated on the basis of classical prediction methods in unweighted networks.In this thesis,we present several novel link prediction methods from the simple unweighted networks to the complex type of networks,i.e.weighted network and multiplex network,and validate the effectiveness of these methods.The main contributions of this thesis are briefly depicted as follows:(1)Drawing on the difference of connectivity strength for different paths between nodes,we propose a novel structural similarity method based on the quasi-local paths in undirected unweighted networks.Generally,the traditional path-based similarity methods only focus on the contribution of the total number and length of paths,which neglects the problem of path heterogeneity for paths with same length.In order to quantify the path weight between nodes,from the view of intermediate links and nodes on paths,we consider the influence of the link degree of existing links and the connectivity influence of intermediate nodes,respectively.Finally,a Local Weighted Path(LWP)index is proposed to differentiate the contributions between paths.The experimental results show that LWP index outperforms other seven traditional prediction baselines and the performance of LWP index is inversely proportional to the link degree of intermediate links.(2)In order to further address the path heterogeneity problem in unweighted networks,considering the strength of interactions among nodes on different paths,we propose a novel structural similarity-based method that depends on the interactions among paths.Most existing path-dependent methods only pay attention to the contributions of paths with specific length,which neglects the interactions of paths with different length for performance improvement.It is well known that short paths generally have a better effect on performance improvement than long paths.Therefore,we simulate the interactions among different paths by the mechanism of resource-traffic flow on networks.Furthermore,the more the intermediate nodes of a path receive resources from nodes on short paths,the greater the contribution of this path is.Finally,a path-dependent link predictor based on the Resource receiving process from Short Paths(RSP)is proposed.The experimental results show that RSP index always has better prediction performance than other nine traditional methods.(3)In weighted networks,considering that the topological structure of links can reflect the strength of links to some extent,we propose a weighted prediction method based on the coupling weight of links.Generally,the weight of links is naturally applied to the study of link prediction in weighted networks,which always neglects the influence of the topological structure on prediction performance.Therefore,we regard the function of link degree as the topological weight of links in weighted networks,and define the concept of coupling weight of links that combines the contribution of the natural weight and topological weight.Finally,a weighted prediction method is proposed based on the coupling weight of links.The experimental results show that the coupling weight-based method has better performance than the methods that rely completely on the natural weight or topological weight of links.(4)From the view of multiplex networks,utilizing the dependency of interlayer structure,we propose a prediction method based on the layer relevance of multiplex networks.Compared to the single-layer networks,multiplex networks provide more diverse structural features.However,the existing link prediction methods primarily focus on the internal structural information derived from single-layer networks and the influence of interlayer information on prediction performance is paid less attention.With the aid of different measures of layer relevance,we examine the contribution of the information of interlayer structure on prediction performance in multiplex networks.Considered the global overlap rate and the Pearson correlation coefficient between layers,a Node Similarity Index based on Layer Relevance(NSILR)that combines the intralayer and interlayer information is proposed.The experimental results show that our NSILR index can significantly improve the prediction performance compared with the traditional link prediction methods that are based only on the single-layer networks.Furthermore,the NSILR index can well solve the cold-start problem of link prediction.Due to the influence of different layer relevance measures on prediction performance,we propose the entropy of layer relevance to prejudge the contribution of interlayer information from multiplex networks. |
PDF Full Text Request