基于因果推断的图注意力网络

Graph Attention Networks Based on Causal Inference

图注意力网络(Graph Attention Network,GAT)是一种重要的图神经网络,在分类任务中有着广泛的应用。但是当图中邻域节点受到干扰时,模型分类准确度会受影响而降低。对此,提出一种基于因果推断的图注意力网络(Causal Graph Attention Network,C-GAT)以提升网络的鲁棒性。该模型首先计算目标节点的邻域与其标签之间的因果权重,并以此对邻域进行因果采样;然后计算采样后邻域与目标节点之间的注意力系数;最后根据注意力系数对邻域信息进行加权聚合,获得目标节点的嵌入特征。在Cora,Pubmed和Citeseer数据集上的实验结果表明,在无扰动的情况下,C-GAT的分类性能与经典模型持平;在有扰动的情况下,相比于经典模型,C-GAT的分类准确度至少提升了6%,在鲁棒性和时间成本上有着较好的平衡。

Graph attention network(GAT)is an important graph neural network with a wide range of applications in classification tasks.However,when the neighborhood nodes in the graph are disturbed,the model classification accuracy will be affected and degraded.In response,a graph attention network based on causal inference named causal graph attention network(C-GAT)is proposed to improve the robustness of the network.The model first calculates the causal weights between the neighborhood of the target node and its label and uses them to sample the neighborhood.Then the attention coefficient between the sampled neighborhood and the target node is calculated.Finally,the embedding features of the target nodes are obtained by weighted aggregation of the neighborhood information based on the attention coefficients.Experimental results on the Cora,Pubmed and Citeseer datasets show that the classification performance of C-GAT is on par with the classical model in the case of no perturbation.In the presence of perturbations,the classification accuracy of C-GAT improves by at least 6%compared to the classical model,with a better balance of robustness and time cost.