Abstract: In view of the problems that traditional network representation learning method could not learn the correlation of network structure of nodes, we proposed a learning model for network structure representation based on neighborhood information. Firstly, the model defined a calculation method of structural similarity between nodes based on neighborhood information, and modeled the structural similarity between nodes in different neighborhood ranges. Secondly, the deep autoencoder was constructed, the node structure similarity was used as supervised information to optimize the network representation, and the node structure information was learned
 in the process of network embedding. Compared with three related algorithms such as node2vec,SDNE,struc2vec, the experimental results show that the proposed method has better ability of identifying the network structure, can learn the structural correlations between nodes, and the obtained network representation can be suitable for role identification related tasks. In addition, the experimental results of cross-network classification also demonstrate the potential of this method in transfer learning.

Key words: structural identification, network representation learning, network analysis, auto encoder, role identification