Network Embedding Method Based on Semantic Information

Graph embedding is the resultful method to map the graph in the low-dimensional vector space. Now most existing embedding methods to learn nodes representations mainly focus on obtaining nodes adjacent and feature information, but they ignore the state that there is also semantic information between nodes. Therefore, it is proposed a graph embedding method, which introduces point mutual information to compute the semantic similarity between nodes, the basic idea is to count the probability of two nodes appearing simultaneously in a sentence. And it learns representations by modeling the sum of the squares of the difference between point-wise mutual information and the inner product of node vectors, and theoretically shows that using point mutual information can also obtain a log-linear relationship between graph topological by leveraging the invariance property of difference between nodes. Finally, the study selects 5 social networks datasets for node classifications, clustering tasks, and compares them with 6 graph embedding methods and 4 methods based on graph neural network, the result demonstrates that the direct method does not negatively impact accuracy on many downstream applications, and outperforms all the baseline methods. In addition, the computation complexity of our method is lower than the worst-case.