Research On Attributed Network Embedding Methods Based On Matrix Factorization

Network embedding aims to learn the low-dimensional and continuous vector representations for all nodes in networks,which is used as the input feature for many complex networks analysis tasks.The majority of existing network embedding methods only focus on network topology structure,however,nodes themselves are usually accompanied by rich attributes which imply certain properties of networks in the real world.Therefore,in order to make the learned vector representations of nodes preserve the useful properties of network structure and attributes of nodes simultaneously,how to properly incorporate network structure and attributes information of nodes into an unified network embedding framework is a promising research subject.Attributed network embedding is defined as network representation learning with attributes information of nodes in this paper.There are mainly two challenges in the study of attributed network embedding.1)Since network structure and node attributes are from two kinds of sources reflecting the features of the same network from different views,how to extract the informative features from them at the same time to complement each other is a key issue.2)In the era of big data,the number of nodes is huge in real networks,which raises higher request to the scalability of attributed network embedding.To handle the aforementioned challenges,we do the following work.First,to deal with the text attributes of nodes,under the framework of non-negative matrix factorization,the Text Enhanced Network Embedding(TENE)method is proposed.By discovering the consistent relation between text cluster structure based on text attributes of nodes and node representations,the learned network embedding gets more information and identifiability.Under the guidance of proximity matrix based on network structure and cluster membership matrix derived from clustering for text attributes,TENE integrates available characteristics of both network structure and text attributes together into the final node representation.By taking network embedding as input to execute node classification on three real datasets,we prove the superior performance of our proposed method TENE.Second,we propose a more general attributed network embedding method called CCANE(Consistency Constrained Attributed Network Embedding).It models the proximities based on network structure and attribute information respectively under the frame of symmetric matrix factorization,and build the consistency and complementarity relation between the two types of node representation,then concatenate them as the final node representation.Besides,CCANE breaks up the model and optimization into many easy and mutually independent sub-problems,resulting the original problem is able to be solved in parallel.By executing tasks of node classification and visualization,we prove the effectiveness and efficiency of our proposed method.