Node Classification Using Graph Convolutional Network with Dropout Regularization

In this paper, node classification using graph convolutional network (GCN) is studied. First, problem formulation of node classification is described. Then, the graph convolutional operator (GCO) is constructed and it is combined with nonlinear activation function to obtain the two-layer GCN for tackling the node classification problem. Next, the dropout regularization is incorporated into the GCN for solving the overfitting problem. Because input feature data is very sparse, sparse dropout is used in the first layer and general dropout is employed in the second layer. Finally, citation network datasets, t-SNE data visualization, ablation study, and classification accuracy are used to evaluate the performance of the GCN with dropout regularization.