Research On Node Classification Based On Data Augmentation Graph Convolutional Network

Node classification is an important research topic in graph learning.The graphical neural networks(GNN)has achieved the most advanced performance in node classification.However,existing GNN solve the problem of balancing node samples of different classes.For many real-world scenarios,some classes may have much fewer samples than others,in which case GNN training will underestimate the samples of these few classes,resulting in suboptimal performance.Here we propose graphic neural network based on unbalanced node classification(UNCGNN),which is a GNN method for unbalanced data node classification.Specifically,our model uses an oversampling algorithm to synthesize a few class samples,which is stable and efficient.We encode the similarity of nodes in the embedded space to ensure high-quality sample generation.At the same time,we train edge generators to build relational information to generate samples.Finally,the adaptive tag propagator adaptively assigns pseudo tags to generate samples,and the node balance graph is used for the final classification task.Experimental results on public data sets show that this method outperforms the existing baselines.(1)To solve the problem of poor generalization performance caused by less data for deep learning models.Data augmentation for non-Euclidean graphs is studied.In the context of improving the node embedding method of semi-supervised node classification.And the two-way data augmentation graph convolutional networks(TDA-GCN)is proposed.A subgraph augmentation method is designed to add or delete adjacent edges of each node on the subgraph,the one with the smallest change in graph entropy is used as an augmented graph,and then the embeddings are extracted from the topological structure and node attributes of the augmented graph and the original graph,and finally the attention mechanism is used to adaptively fuse of learned embeddings.Experiments have proved that this method is superior to the currently popular methods and has achieved good results.(2)Graph Neural Networks(GNN)learn node representations of graphs through topological structure neighborhood propagation and aggregation.It has achieved great success in graph learning tasks and is widely used in node classification tasks.However,graphs in the real world are usually sparse and noisy,and noise can also affect surrounding nodes through topological structure,which negatively affects the performance of GNN.In this thesis,we propose a new framework to deal with these problems.Specifically,first,we design an edge generator to densify the sparse graph using node similarity and reduce the weight of noisy edges.Then,adaptive label propagator generates pseudo-labels based on topological structure,and the dynamic and adaptive weighting strategy of the adaptive label propagator can overcome the shortcomings of GNN on noisy labels.Finally,we use the topology information and node information of the obtained new graph for final classification and pseudo-label reweighting through GNN.Experimental results on public datasets indicate the superiority of our method over existing baselines,demonstrating the robustness of our method to sparse and noisy graphs.(3)Node classification is an important research topic in graph learning.Graph neural networks(GNN)have achieved the most advanced performance in node classification.However,existing GNN solve the problem of the balance of node samples of different classes.And for many real-world scenarios,some classes may have far fewer samples than others,training of the GNN in this case would underestimate samples of these minority classes and result in suboptimal performance.Here we propose graph neural networks based on unbalanced node classification(UNCGNN),a GNN method for node classification of unbalanced data.Specifically,our model uses an over-sampling algorithm to synthesize minority classes samples and this method is stable and effective.We encode the similarity of nodes in an embedded space to ensure a high quality sample generation.At the same time,we train the edge generator to build the relationship information to generate samples.Finally,the adaptive label propagator adaptively assigns pseudo label to generate samples,and the node balance graph is used for the final classification task.Experimental results on public datasets indicate the superiority of our method over existing baselines.