Semi-supervised Learning With Dual Channel Graph Convolutional Networks

Graph data widely exists in real life,such as literature citation relationships,social networks and so on.Since it can naturally describe the relationship between data,it is of great significance and value to study graph data.Traditional graph analysis methods are mostly based on the statistical information of graph data or manually designing features.However,it performs inefficiently and the process of designing features can be time-consuming in most cases.Recently,deep learning has been widely used in many research fields due to its powerful feature representation ability and no need for excessive prior knowledge.To some extent,the need for statistical information and manually designing features has been reduced.Therefore,using deep neural networks to process graphs has drawn lots of research interests.This leads to the concept of graph neural networks.Graph neural networks can be divided into two types based on the definition of graph convolution,i.e.,spectral domain based,and spatial domain based.Graph Convolution Network(GCN)can be regarded as a model based on the spectral domain and the spatial domain.GCN is well studied because of its solid theoretical basis and concise forward propagation rule.GCN generalizes convolution operations from traditional grid data to graph data in the non-Euclidean domain and achieves good performance on related tasks.However,training a good GCN often requires a lot of labels as well as validation set,while in real life,the acquisition of label data is often time-consuming and expensive.Therefore,how to train a GCN with a small amount of label data,i.e.,low label rate,has become a new research problem.In this paper,we use dual channel model to overcome the problem of GCN at low label rates.The dual channel model consists of two graph convolution networks with same structure but different parameters.GCN performs well in label transduction propagation and has high confidence in the prediction of some unlabeled samples.We use high-confident predictions as pseudo labels and expand the label data set when the label data is very limited.However,since pseudo labels are obtained through the prediction of model,there may be some noisy labels.Trusting all the pseudo-labels will cause propagation errors.A recent study on noisy label learning found that the neural networks first fit the clean data,and then fit the noisy data.We use the same method: first,our dual channel model is trained using clean data to get a good initialization,and then select pseudo labels with high confidence to expand the label set.During the training process,we gradually increase the amount of pseudo-labeled data.The specific selection of pseudo labels is: first select the prediction samples of two models with high confidence generated by the softmax function;then select the samples with same predictive labels by our dual channel model among these high-confident samples and add them to the label set to expand the label set.The comparison of node classification results with other graph learning methods demonstrates the effectiveness of our proposed method,i.e.,our method can achieve good performance at a very low label rate.