A Semi-Supervised Graph Neural Network with Confidence Discrimination

Existing graph neural network methods usually depend on a large amount of labeled data, but labeled data is often scarce in the real world. In the case of less labeled data, utilizing correct pseudo-labels for model training can improve the model performance effectively. However, existing pseudo-labeling methods often use a fixed confidence threshold for all classes, leading to class imbalance and low data utilization. To solve this problem, a semi-supervised graph neural network method based on confidence discrimination is proposed, which can make full use of unlabeled data to facilitate semi-supervised node classification. Our method considers the learning state and difficulty of different classes of nodes and designs an adaptive confidence discrimination module. It assigns different confidence thresholds to each class of node, and uses unlabeled nodes with high confidence to expand the label set continuously. Our method can learn more discriminative node features to improve the model performance. On five publicly available datasets, the accuracy of the proposed method is improved by 2.1% on average compared with other methods, and in particular by 5.7% on the Flickr dataset. Extensive experiments verify the effectiveness of the proposed method.