Figure 5:

Out-of-sample classification accuracy as a function of the difference between classes for different graph embedding methods. Graphs in the sample are distributed as a multilayer SBM with n = 256 vertices, two communities, four different connectivity matrices that correspond to the class labels, and a training and test samples with m = 40 graphs. As the separation between the classes increase, all methods improve performance, but MASE and OMNI show the most gains.