Discriminative Graph Convolutional Networks for Semi-supervised Node Classification

Abstract

Graph Convolutional Networks (GCNs) gain remarkable success in graph-based semi-supervised node classification task. Despite their success, most GCNs still exist several challenges. For this task, it is necessary to draw nodes in the same class close and push ones from different classes apart. However, GCNs smooth the node's representation by aggregating information within node neighborhoods, despite whether the connected nodes are from the same class or not. The smooth property overlooks the intra-class similarity and inter-class diversity, which leads to GCNs failing especially on low homophilic or heterophilic graphs where most nodes have neighbors from different classes. In this paper, we propose the Discriminative Graph Convolution Networks (DGCN), an extension of GCN model with discriminant modules: intra-class smoothness and inter-class sharpness. The modules effectively strengthen the intra-class similarity and the inter-class differences by introducing available label information into the convolution process. Extensive experiments on six benchmark datasets demonstrate the effectiveness of DGCN in semi-supervised node classification. Code available at https://github.com/AIG22/DGCN.