Robustness Research On Direction Of Feature Propagation And Depth Of Feature Propagation Of Graph Convolutional Neural Network

Due to its superior performance in semi-supervised learning tasks,convolutional graph neural networks(GCNs)and their derivatives have gotten a lot of interest in recent years.Many domains,including computer vision,natural language processing,social networks,and applied chemistry,have benefited from convolutional graph neural networks.However,most current methods have two flaws: the first is that the algorithm is not robust to the direction of feature propagation,because most existing methods are based on eigenvalue decomposition and cannot be used for directed graphs;and only a small portion of the performance of methods that can be used in directed graphs is higher than that of undirected graphs;and the second is that most networks are not robust enough for the depth of feature propagation.Most approaches can only conduct shallow learning at the moment.Too much or too deep a network will result in feature over-smoothing,making it unable to fully exploit the full-image information to increase learning efficiency.As a result,the robustness of the feature propagation direction and the robustness of the feature propagation depth are investigated in depth in this research.The following is the main research project:(1)At the moment,graph convolutional neural network datasets are mostly undirected graphs.As a result,in this paper,a directed citation network dataset called CCFcitation is devised and established.The data is organized as a directed citation network with nodes representing authors.There are 6817 nodes in total,each with its own author category.When two authors have a paper citation link,a directed edge is added between the two nodes,giving a total of42662 edges.(2)This research offers a robust directed convolutional graph neural network algorithm based on stationary distribution to address the resilience of feature propagation direction.The algorithm’s basic steps are as follows: To begin,complete the original directed graph’s relationships,anticipate the prospective relationships that are missing from the graph,and create a more accurate directed graph.The graph’s approximate stationary distribution is then determined,and a stationary directed Laplacian propagation matrix is obtained.Finally,to finish the semi-supervised node classification task,feature propagation is conducted and a classifier is trained.The difference in each index between the directed and undirected versions of the four datasets of Cora,Citeseer,Pubmed,and CCFcitation of the technique suggested in this research is less than 2%,and it outperforms previous methods.(3)This research offers a sequential convolutional graph neural network technique based on Transformer to improve the robustness of feature propagation depth.The algorithm’s basic step is to extract features after each convolution propagation,sequence them,and send them to the modified Transformer encoder to learn a global feature,which is then used for classification.Deep networks’ feature over-smoothing problem is avoided.The accuracy of the technique described in this research on the three datasets Cora,Pubmed,and Citeseer consistently increases by roughly 10% as the network depth increases from 1 to 25,and there is no performance degradation as the network deepens.