Study Of Graph Neural Network Based On Causal Sampling

The research of graph neural network has become one of the hotspots in the field of deep learning.The classification robustness of graph neural network has always been the frontier topic of graph neural network research.It is of great significance to improve the classification performance and scalability of graph neural network model.Aiming at the problem of classification robustness of graph neural network,this paper proposes the research of graph neural network based on causal sampling from the perspective of causal inference.The specific work is as follows:Firstly,a causal sampling method on graph data is proposed.Establish the causal model among nodes,labels and perturbation in the graph data,cut off the back door path in the causal model through the back door criterion,and obtain the intervention formula on the model.The kernel probability density estimation function is used to estimate the intervention formula to obtain the causal weight of the nodes in the graph relative to the label.The non-uniform sampling of nodes on the graph is realized by causal weight.Secondly,according to the characteristics of GraphSAGE model sampling first and then aggregation,a Causal GraphSAGE model based on causal sampling is proposed.By analyzing the causal confusion caused by random sampling and external perturbation in GraphSAGE model,thedirected acyclic causal model on GraphSAGE is established.The back door criterion is used in this model,and causal sampling is introduced into the sampling process of GraphSAGE to sample the causal node set.The information in this causal node set is aggregated by training a set of aggregation functions,and then the embedded representation of the nodes is applied to the classification task.In the experiment,in the case of perturbation,the classification accuracy of Causal GraphSAGE on Cora,Pubmed and Citeseer datasets is 6.2%,7.5% and 5.9% higher than that of GraphSAGE model.Finally,according to the characteristics of Graph Attention Network aggregating neighbors through attention coefficients,the Causal Graph Attention Network model based on causal sampling is proposed.By analyzing the internal causal problem of Graph Attention Network model,combined with the causal confusion caused by the damage of attention coefficient caused by external disturbance,a directed acyclic causal model on Graph Attention Network is established.The back door criterion is used in this model,and the causal sampling mechanism is introduced into the attention layer of Graph Attention Network to obtain the set of causal attention neighborhoods through causal sampling.In the perturbation experiment on Cora,Pubmed and Citeseer datasets,the classification accuracy of Causal Graph Attention Network model is 6.5%,6.8% and 6.0% higher than that of Graph Attention Network model.