Semi-supervised Graph Learning with Few Labeled Nodes

Abstract

Graph-based semi-supervised learning, utilizing both a few labeled nodes and massive unlabeled nodes, has aroused extensive attention in the research community. However, for the graph with few labeled nodes, the performance of Graph Convolutional Networks (GCNs) will suffer from a catastrophic decline due to its intrinsic shallow architecture limitation and insufficient supervision signals. To accommodate this issue, we propose a novel Self-Training model (ST-LPGCN) which reinforces the pseudo label generation on the GCNs with Label Propagation algorithm (LPA). By making full use of the advantages of GCNs in aggregating the local node features and LPA in propagating the global label information, our ST-LPGCN improves the generalization performance of GCNs with few labeled nodes. Specifically, we design a pseudo label generator to pick out the nodes assigned with the same pseudo labels by GCN and LPA, and add them to the labeled data for the next self-training process. To reduce the error propagation of labels, we optimize the transition probability between nodes in LPA under the supervision of the pseudo labels. The extensive experimental results on four real-world datasets validate the superiority of ST-LPGCN for the node classification task with few labeled nodes.