Semisupervised Classification with High-Order Graph Learning Attention Neural Network

Figure 2

Comparison of high-level attention mechanism and GAT attention mechanism. Node A is the node that we need to request a hidden representation. (a) shows the attention mechanism of HGAT. It not only refers to the 1 hop neighbors but also calculates the attention coefficients of higher-order neighbors, and as the order increases, the color gradually becomes lighter; that is, the attention coefficient is distributed less. (b) shows the attention coefficient calculation method of GAT, and the attention coefficient is obtained only by calculating the neighbors of 1 hop.

(a)

(b)