Sequence to Sequence Network for Learning Network Representation

Abstract

Network representation learning is an important way for learning the low dimensional vector of nodes in the network, with preserving certain structural information between nodes in the original graph. Most existing network embedding models use truncated random walks and shallow architectures which do not fully obtain the nonlinear information and neighborhood information of the network. In this article, we propose a novel method for network representation learning which generates low-dimensional representation vectors for each node in the graph by obtaining the local and global structure information of the network. Unlike previous work, we use the hybrid BFS and DFS methods to sample the neighbor information of each node instead of using the uniform sampling method in DeepWalk to generate the linear sequences. After obtaining the linear sequences, we use a sequence to sequence network that contains a teaching sequence which is proved effective in capturing the nonlinear information of graph, to learn the reconstruction error of the input sequence and the output sequence. We named our method SSNR which is not only preserve both the local and global network structure information, but also capture the nonlinear information from network to achieve more discriminative node representation. To verify the effectiveness of SSNR, we employ the learned node representation as features in downstream experiments with node classification and graph visualization tasks. The experimental results of different datasets demonstrate that SSNR outperforms many state-of-the-art baseline models in these tasks.