Generative Adversarial Network-assisted Graph Convolutional Network (GANA-GCN)

Abstract

Generative adversarial networks (GANs) have attracted considerable attention due to the generator having fewer restrictions compared to existing techniques like Boltzmann machines. However, in the field of graph convolutional networks (GCNs), most efforts to improve the embedding quality have focused on making node representations contain more structural information, and a method to optimize the node embedding using the adversary concept has thus far been lacking. In the present work, we discover that the adjacency matrix generated from node labels is helpful to learn better node embeddings and improve semi-supervised node classification accuracy. We thus propose GANA-GCN, which consists of a generator <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$G$</tex> that learns node embeddings and labels to make a classification and a discriminator <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$D$</tex> that encourages node embeddings include label information when the discriminability is weakened. Experimental results demonstrate that GANA-GCN is effective to improve the classification accuracy on a non-Euclidean dataset. Sensitivity analysis further illustrates the effect of key parameters on the overall performance.