Semisupervised Classification with High-Order Graph Learning Attention Neural Network

Figure 4

Classification accuracy of ablation experiments on Cora datasets with different edge retention rates. On the Cora dataset with different edge retention rates, the classification accuracy of HGLAT and the three No-Reg, No-Fus, and GLAT ablation experiments are compared, where No-Reg means that the graph structure obtained by the graph learning module is not symmetrized and regularized constraints; No-Fus means that the decoder graph structure and the original graph structure are not allowed to merge; that is, formula (10) is not used. GLAT means that the edges are reweighted only by the relationship between the first-order neighbors.