A Hypergraph Convolutional Neural Network for Node Classification of Citation Network Data

This paper provides a hypergraph convolutional neural network (HGCNN) for node classification of irregular citation network data. First, hypergraph representation is used to describe the more general relation among the data, and the node classification problem is described. Then, the hypergraph convolutional operator and activation function are employed to construct a two-layer HGCNN model with sparse dropout regularization to transform the node feature vector to predicted label. Next, the cross entropy between true label and predicted label are chosen as the loss function to train the model weights and bias vector of HGCNN model. Finally, three citation network datasets are applied to demonstrate the effectiveness of the HGCNN model according to classification accuracy, learning curve and data visualization.