GNN-LS: An Information-Enhanced Method for Label Scarcity on Graphs

Abstract

Label scarcity severely hinders the performance of the node classification task on graphs. To overcome the lack of supervised information provided by labels, some work has attempted to attach a pretext task based on self-supervised learning, aiming to mine intrinsic information from the graph data itself as a complement. However, the representation learning capacity of existing methods is undermined by two main factors: 1) insufficient supervised information caused by the underutilization of limited labels, and 2) the multi-view expressiveness capabilities of intrinsic information for an individual model are restricted. Therefore, we propose an information-enhanced method for label scarcity on graphs (GNN-LS). For supervised information, we develop a dual guidance strategy that effectively utilizes unlabeled nodes while considering prediction reliability to provide stronger extra supervised guidance for the model. Meanwhile, considering the possible redundancy of the model in acquiring two types of information, we construct a regularization to reduce it. Finally, to effectively expand the range of views on the intrinsic information mined, we design a multi-view integration mechanism, which determines coefficients between intrinsic information from different views according to their respective fitness for the current dataset and label rate. Comprehensive experiments show that our method achieves significant performance gains compared with its peers on seven benchmark datasets.