Community-Specific Deep Learning Model For Embedding On Attribute Networks

This paper mainly studies the community-specific deep learning model for embedding attribute network.That is,integrating the community information into the deep learning model for embedding attribute network.To achieve this goal,we need to solve three sub-problems: first is how to better integrate attribute information with the network topologies;second is how to generate community information in the representation model;the last is how to use the community information in the representation model.To achieve this goal,we have proposed two network embedding methods.1.The first method is a community-specific attribute network embedding method based on Variational Auto-Encoder(VAE).This approach proposes a ”sharedindependent” framework to deal with the relationship between network topology and node attributes.For the encoding phase where these two kinds of information have commonality,the shared encoder is used to derive latent embedding.For the decoding phase,a dual decoder structure is designed,which is consist of two different generation mechanisms.In particular,the topology decoder uses a new topology generation mechanism,which considers the generation of link between nodes from the perspective of the community.We then formalize this generation mechanism as the form of neural network.In this way,the community information can be integrated into the whole model.At the same time,in order to generate community information,a Gaussian Mixture Model is introduced to connect the encoder and the decoder,so that the embedding derived by the encoder can be converted to community information and provided to the decoder for reconstruction.This part can also be formalized as a neural network and integrated into the model.We conducted experiments on real networks.The results show that our method is superior to the comparison method,which proves the validity of our model.This will help further improve the quality of embedding to help better apply to network analysis tasks.2.The second methods is a community-specific attribute network embedding method based on Graph Convolution Network(GCN).This approach is a unified framework which combines Graph Convolution Network with community detection.First,the Graph Convolution Network is used as the encoder to integrate the network topology with node attribute.Second,in the community detection module,a loss item based on the modularity is designed to maximize the modularity of network along with minimizing the loss of the model.In this way,an accurate community information can be generated.Finally,a dual encoder structure and a network reconstruction module are designed.The former generates two different community detection results.After that,a relative entropy(KL)minimization process between this two results is employed,in which process integrates community information into the whole model.The latter reconstructs the network topology using an inner product operation to further constrain the parameter of GCN.These two modules work together to allow the whole model work in an unsupervised way.We conduct experimental on the real network.The experimental results further confirm the validity of our model.The improvement in the quality of the embedding will better help network analysis tasks.