Semisupervised Classification with High-Order Graph Learning Attention Neural Network

Figure 1

HGLAT model architecture. The dashed box is a data set without a graph structure. The initial graph structure needs to be generated through kNN clustering to be used as the input part of IVGAE. This step is not required for a data set with a graph structure. IVGAE is a graph learning module, which generates a new graph structure , and calculates the graph learning loss. is symmetrized and combined with the initial graph structure to obtain , and then, is used as the input of HGAT. HGAT calculates high-order neighbor attention coefficients, performs semisupervised node classification, and calculates the classification loss. The two modules are jointly optimized to obtain semisupervised classification results.