Research On Graph Contrastive Learning Problem In Node Classification Tas

Graph neural networks are widely used in processing graph data and node classification tasks,and graph contrastive learning is an effective self-supervised learning method widely developed in the field of data mining.Due to the high complexity of graphs and the high difficulty of obtaining node labels,the task of obtaining node labels for node classification in datasets with few labels becomes a challenge.The paper is a study of graph contrastive learning in the node classification task,carried out under self-supervised learning by improving the part of data augmentation in order to improve the accuracy of node classification.Based on this,the paper does the following three things:In the first part,a graph contrastive learning model based on negative-sample-free loss and adaptive augmentation is proposed to address the problems of random enhancement of the input graph and the need to construct losses using negative samples in graph contrastive learning methods.In the model,the centrality of the node degree in nodes and edges by random enhancement and thus improve the robustness of the model.The embedding correlation-based loss function which does not rely on non-symmetric neural network architectures is used to guide the model learning.Negative samples are not required in this loss function,avoiding the problem of negative samples that are difficult to define in the graph and negative samples increase the computational and storage burden of constructing losses.Node classification experiments on the citation dataset show that it outperforms many baseline methods in terms of classification accuracy.In the second part,a graph contrastive learning model based on graph diffusion is proposed to address the problem that most existing methods use uniform data enhancement schemes,such as uniform deleted edges and uniform transformation features leading to poor performance.The model performs graph diffusion and downsampling on the input graph,represented by an identically weighted encoder network and MLP learning nodes and the use of a contrastive loss function achieves high node classification accuracy,and node classification experiments on citation datasets show that comparison with the baseline approach validates the effectiveness of the model.In the third part,a graph contrastive learning model based on hierarchical graph convolutional networks is proposed to address the problem that the shallowness of most neighbourhood aggregation-based models resulting in models from obtaining sufficient global information.The model aggregates nodes with similar structure into newly generated nodes,and after coarsening the graph to restore each node in the graph to its previous representation,the graph encoder part is placed together with the data enhancement part.Lastly the model introduces a contrastive loss function to train the model.The effects of node classification experiments on citation datasets show it can improve the classification accuracy of the model better than many baseline methods.