Uncertainty-Confidence Fused Pseudo-labeling for Graph Neural Networks

Graph Neural Networks (GNNs) have achieved promising performance for semi-supervised graph learning. However, the training of GNNs usually heavily relies on a large number of labeled nodes in a graph. When the labeled data are scarce, GNNs easily over-fit rto the few labeled samples, resulting in the degenerating performance. To address this issue, we propose a novel graph pseudo-labeling framework. The proposed framework combines both the predication confidence and approximate Bayesian uncertainty of GNNs, resulting in a metric for generating more reliable and balanced pseudo-labeled nodes in graph. Furthermore, an iterative re-training strategy is employed on the extended training label set (including original labeled and pseudo-labeled nodes) to train a more generalized GNNs. Extensive experiments on benchmark graph datasets demonstrate that the proposed pseudo-labeling framework can enhance node classification performance of two alternative GNNs models by a considerable margin, specifically when labeled data are scarce.