Abstract
Graph Convolutional Networks (GCNs) play a vital role in graph learning tasks such as semi-supervised learning. However, a GCN model requires a large amount of labeled data for verification and model selection, and learning on sparse labeled graphs is still a challenging issue. In order to solve this problem, this paper propose an auxiliary learning task enhanced graph convolutional network (A-GCN), which combines the target supervised learning task of the GCN model with the auxiliary unsupervised learning task to correct its network's learning. The experimental results demonstrate that A-GCN can achieve a significant performance improvement compared with state-of-the-art methods on a weakly supervised graph.