Research And Application Of Bipartite Network Representation Learning

With the rapid development of the internet,complex networks have been widely concerned in many research fields in recent years,such as recommendation,urban risk assessment,and criminal behavior prediction.The bipartite network exists widely in these fields,and network representation learning is an effective network analysis method,aiming to map the representation of network nodes to low-dimensional vector space,but most of the current network representation learning methods are based on homogeneous networks,without considering the special properties of bipartite networks.Therefore,this paper proposes two network representation learning models of bipartite networks according to the implicit relationship and topological properties of the bipartite network.The specific work is as follows:Firstly,a deep learning model,BiNE-IEI,is proposed,which integrates the implicit relations of bipartite networks.This model firstly obtains the implicit relation between nodes of the same type through mapping method and then fuses the implicit relation and explicit relation(explicit link)in the network through the deep learning framework auto-encoder(AE)to learn the representation of nodes.Finally,the highly non-linear structure of the network is also preserved.Secondly,a kind of deep learning model BiVAE based on variational encoder(VAE)is proposed,which can preserve the topological structure of the bipartite network.A node-level triplet is designed to constrain the relationship between two different types of nodes in the network,that is the similarity between neighborhood nodes is greater than that between non-neighborhood nodes to improve the representation learning effect of bipartite network nodes.BiVAE can not only model the uncertainty of nodes in the network through Gaussian distribution but also preserve the local and global structure of the network.Finally,based on BiVAE model,an empirical analysis is conducted on 110 alarm data of a city.The possibility of alarming behavior is closely related to the alarming person and the place of the crime.Based on 110 alarm data,this paper constructed a bipartite network between the person and the place to analyze the alarming behavior of the person and the place and proposed two kinds of urban area risk calculation indexes to describe the possibility of alarming behavior of the place.Based on the theory of complex network,this paper improves the existing learning method of bipartite network representation learning and proposes two improved deep learning models of the bipartite network.At the same time,based on the model BiVAE,the empirical analysis of 110 alarm data is carried out.Mining more information about the possibility of alarm behavior in urban areas will help to improve the ability of prediction and prevention of urban regional risk,which has important application value.