Abstract: 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.

Key words: Graph attention networks, Causal inference, Attention mechanism, Causal weight