General Structure Preserving Network Representation Learning

As one of the common research objects in the field of data mining,network is ubiquitous in our lives.Many real relationships can be abstracted into networks,where the nodes represent data objects and the edges represent relationships between data objects.With the continuous development of information age.data scale grows rapidly,how to use low-dimensional dense vectors to represent large-scale sparse networks is a key problem in data storage and processing.The problem has gradually become the core of researchers' attentionIn recent years,the related work of network representation learning has made great progress,and various types of methods are constantly emerging.From the perspective of representing network,there are methods based on eigenvector manifold invariance,methods based on neural network learning,and methods based on non-negative matrix factorization.From the perspective of preserving property of networks,there are methods that preserve micro-structure,such as node neighbors,network triangle structure,and also preserve meso-structure,such as community structure of networks.However,in many meso-structures of networks,besides community structure,there are other kinds of general structures such as binary structure,star structure,core-peripheral structure,mixed structure,etc.Therefore,how to preserve general structure and propose an effective network representing learning method deserves deep research.Moreover,there is not only global topological relationship in a network,but also abundant attribute information along with nodes.How to make the results of network representation have the ability to preserve the similarity of structure and attribute deserves our farther attention.In view of the above problems,the research of this work mainly includes:(1)We propose a network representation learning method GS-NMF which is based on non-negative matrix factorization and capable of preserving general structure of networks.In this work,the constraint of maximize community modularity in M-NMF is adjusted to three-factor factorization model,since the three-factor factorization model has the ability to learn general structure of networks.Therefore,our proposed method has advantage of overcoming the weakness of original method.We do experiments on artificial and real networks with various general structures including community structure,binary structure,mixed structure,then verify the effectiveness of GS-NMF in clustering,classification and visualization.(2)Based on a network representation learning model which preserves general structure,we integrate attribute information and propose a general structure preserving attributed network embedding method GS-ATTR.On the basis of GS-NMF model,this method adds invariance constraint on similarity of attribute during the embedding process.By compared with several existing attribute network representation learning methods on datasets with different structures including commnity structure and mixed structure,our method not only preserves meso-general structure,but also get good results on classification and clustering tasks.Compared with the original GS-NMF model,the proposed method gets a significant improvement,which further confirms that the integration of node attributes helps to improve the effectiveness of analysis tasks.