Research On Attribute Network Representation Learning For Border Node Discovery

The main task of network representation learning is to learn low-dimensional dense vector representation for nodes in the network,and provide efficient data support for network analysis tasks such as node classification,link prediction,node clustering,etc.The existing network representation learning methods mainly include structure representation learning and attribute representation learning.However,researchers usually only consider the normal node characteristics when modeling the structure and attribute characteristics of the network,ignoring the importance of border nodes in the network.Although the number of border nodes in the network is often lower than that of normal nodes,it plays an important role in solving network analysis problems.It is very important to study the representation learning method for border nodes to improve the effectiveness of various tasks in network analysis.At the same time,it also faces many challenges: 1)how to effectively integrate the structural information and attribute information of nodes,and learn more comprehensive characteristics of nodes from multiple perspectives;2)how to model the relationship between border nodes and normal nodes to better distinguish the main features of border node and normal node.In order to solve the above problems,this paper proposes an attribute network representation learning method for border node discovery(refer to as ANE-BN),which integrates the structure information of network and the attribute information of node,and models the relationship between the structure representation and attribute representation of the network,so as to achieve the complementarity between the stability of structure features and the efficiency of attribute features.This is the first time that applies structure information and attribute information to the representation learning for border nodes.It not only considers the attribute modeling of border nodes,but also considers how to integrate attribute information and structure information into the node representation learning model.Specifically,first of all,by fusing of high-order network structure information to calculate the similarity of nodes in terms of network structure,and using cosine similarity to calculate the similarity of nodes in terms of node attribute,to provide data support for structure representation and attribute representation learning.Then,the non-negative matrix decomposition method is used to model the representation learning model of nodes in both network structure and attribute,and the correlation matrix is introduced to model the relationship between structure feature and attribute feature,so as to achieve both the advantages of structure stability and attribute efficiency.In addition,in order to make the vector representation contain more important information,we constrain the model parameters by sparse learning.Finally,in order to increase the stability of algorithm convergence and reflect the rationality of model training,we learn the parameters by alternating optimization algorithm.In the experiment,the effectiveness of the proposed ANE-BN algorithm is verified by using the real-world dataset compared with the existing representative algorithms.The experimental results show that the ANE-BN method can improve the effectiveness of border node representation in various network analysis applications on the premise of ensuring the learning effectiveness of normal node representation.