Research On Link Prediction In Bipartite Networks Based On Multi-scale Characteristics

Link prediction in bipartite networks aims to infer the existence of new relation-ships or underlying interactions between two different types of nodes.It can be used to uncover network evolving mechanisms and be widely applied in terrorist organization identification and recommendation system.In general,existing link prediction algorithms in bipartite networks ignore the inherent complexity and diversity of the links generated between two different kinds of nodes.They cannot explain effectively the network for-mation mechanism,and their performance is obviously reduced.The thesis analyzes the uncertainty problem of the link generative mechanisms,and researches link prediction in bipartite networks in the perspective of multi-scale characteristics.The contributions and innovations are summarized as follows:Firstly,focusing on that the local properties in the original bipartite networks cannot be well preserved by the low dimensional mapping methods,we propose the similarity regularized latent feature model(SRNMF)integrating the local similarity mechanism and low-rank decomposition mechanism.SRNMF constructs a similarity matrix encoding the geometrical information of the networks,and then integrates the matrix into latent feature model of link prediction in the form of similarity regularization.SRNMF combines the macro structure and micro information of the network.The performance of our model is overall better and more stable,compared with seventeen widely applied link prediction algorithms in bipartite networkSecondly,focusing on that existing latent feature models cannot capture the nonlin-ear characteristics of link generated in bipartite networks,we propose kernel-based latent feature model framework for link prediction.In specific,we apply the kernel function to embed the network data into a higher dimensional Hilbert space and extract nonlinear high-order information.The transformed kernel matrix can be regarded as block con-straint information at the mesoscopic scale.Furthermore,considering that single kernel function's limitation of capturing nonlinear features,a link prediction model based on kernel function selection is proposed,which significantly improves the performance of link prediction in bipartite networks.Thirdly,focusing on the complexity and diversity of the link generative mechanisms in bipartite networks,we propose a structure perturbation theory based framework with-out assuming any link generative mechanism at the macro scale.It extends structure perturbation theory in quantum mechanics to bipartite network.Therefore,the proposed framework explores the link generative mechanisms of bipartite networks and improves the prediction performance.Moreover,it is also applicable in analyzing the network evo-lution.In summary,the thesis focuses on the uncertainty problem of links generation in bipartite networks,the methods proposed in this thesis significantly improve the perfor-mance of link prediction and extend a new chapter in the study of link prediction at the multi-scale characteristics.