基于节点相似性的图注意力网络表示学习模型

图注意力网络(GAT)通过注意力机制聚合节点的邻居信息以提取节点的结构特征，然而并没有考虑网络中潜在的节点相似性特征。针对以上问题，提出了一种考虑网络中相似节点的网络表示学习方法NSGAN。首先，在节点层面上，通过图注意力机制分别学习相似网络和原始网络的结构特征；其次，在图层面上，将两个网络对应的节点嵌入通过基于图层面的注意力机制聚合在一起，生成节点最终的嵌入表示。在三个数据集上进行节点分类实验，NSGAN比传统的图注意力网络方法的准确率提高了约2%。

Graph attention network representation learning with node similarity

Graph attention network(GAT) aggregates the information of neighbor nodes to extract the structural features of nodes through the attention mechanism, however, it doesn''t consider the potential node similarity features in the network. To address these problems, this paper proposed a network representation learning method that considered similar nodes in the network(NSGAN). Firstly, at the node level, the NSGAN learned the structural features of the similar network and the original network separately through the graph attention mechanism. Secondly, it aggregated the node embeddings corresponding to the two networks together through the graph-based attention mechanism to generate the final embedding representation at the graph level. In the node classification experiments on the three datasets, the NSGAN improves the accuracy by about 2% over the traditional graph attention network approach. It demonstrates the effectiveness of the NSGAN model.