As can be seen from Figure 7, the results of LE and DeepWalk are not ideal, and the points belonging to different categories are mixed with each other. For lines, different clusters are formed. However, in the center, different types of documents are still mixed with each other. For GraRep, the result looks better because the points of the same color are grouped, but the boundaries of each group are not clear. Obviously, the visualization effect of SDNE is the best in population separation and boundary.
In this paper, a deep network structure embedding, namely SDNE, is proposed to realize network embedding. Specifically, in order to capture the high-dimensional nonlinear network structure, we design a semi-supervised depth model with multi-layer nonlinear functions. In order to further solve the problem of network structure maintenance and sparseness, we use both first-order approximation and second-order approximation to represent the local and global characteristics of the network. By optimizing them in the semi-supervised depth model, the learning representation can reflect the local and global characteristics of the network and is robust to sparse networks. We have experimented with multi-label classification and visualization tasks on real network data sets. The experimental results show that our algorithm has been greatly improved compared with the most advanced algorithm at present.