Research On Deep Representation Learning For Attributed Network

With the development of information technology,massive amount of data generated every day.In reality,a large amount of data is organized by graph structure,e.g.,the social networks,the citation networks.Graph mining can obtain social structure information,understand the dissemination mode of information,and effectively improve the performance of basic tasks such as link prediction,community division,and abnormal point detection.However,large-scale sparse information networks greatly increase the difficulty of social network analysis.In recent years,network embedding models are proposed to embed information networks into low-dimensional spaces,in which every vertex is represented as a low-dimensional vector.Then the learned node features can be well applied to variety of applications.In view of the sparseness of real networks and many legitimate links are actually not observed,many works are proposed to use local and global structure of networks for improving network embedding performance.In reality,each node is usually associated with rich attributes.Some attributed network embedding models leveraged the node attributes in these shallow network embedding models to alleviate the data sparsity issue.However,these previous shallow models fail to capture the non-linear deep information embedded in the attributed network,resulting to the sub-optimal embedding results.Considering the sparseness of network data and the complexity of network structure,this paper proposes a Deep Attributed Network Embedding framework named DANE to capture the complex structure and attribute information.Specifically,the framework first adopt a personalized random walk based model to capture the interaction between network structure and node attributes from various degrees of proximity.After that,DANE construct an enhanced matrix representation of the attributed network by summarizing the various degrees of proximity.Then,a deep neural network is designed to exploit the non-linear complex information in the enhanced matrix for network embedding.Thus,the proposed framework could capture the complex attributed network structure by preserving both the various degrees of network structure and node attributes in a unified framework.Finally,empir ical experiments show the effectiveness of our proposed framework on a variety of network embedding based tasks.Considering the superior performance of network representation learning,this paper proposes an Attributed Network Embedding based Link Prediction algorithm,named ANE-LP.The ANE-LP model utilizes two deep autoencoders to extract rich user features from the user social network and the user attribute information,respectively.Finally,the two types of user features learned are merged toget her by a specific aggregation function for further link prediction.The learned two types of user representations are merged together by a specific aggregation function for further link prediction.Finally,experimental results demonstrate the superiority of the ANE-LP model in link prediction tasks.