Enhancing Graph Neural Networks via auxiliary training for semi-supervised node classification

Abstract Graph Neural Networks (GNNs) have been successfully applied to graph analysis tasks. As a canonical task in graph analysis, node classification has achieved promising results with GNNs. However, training GNNs remains challenging when training samples are limited. For the node classification task, only the known node labels are used as supervision information to train GNNs, while the implicit information of pairwise relationships between nodes is neglected. The label-to-label relationship of node pairs is explicit in labeled nodes and not utilized. Moreover, it is found that the link relationship in a graph is correlated with the labels between nodes in some graphs, like information network. Based on that, we can extract and establish labels for pairwise nodes. With the supervision of labeled pairwise nodes, it can force the predicted labels to conform to the observed pairwise relationships and provide some useful information to boost performance. In view of this, we utilize pairwise relationships by introducing the auxiliary task, node pair classification or link prediction, and propose a novel training framework for enhancing GNNs, namely EGNN. Via jointly training node classification with the auxiliary task, our framework achieves higher classification accuracy for general GNNs models with only a little extra computational cost. Moreover, the adaptive dynamic weighting strategy is designed to balance the training pace of different tasks automatically. We conduct extensive experiments, and the evaluation results suggest the superiority of our framework.