Label-Enhanced Graph Neural Network for Semi-Supervised Node Classification

Graph Neural Networks (GNNs) have been widely applied in the semi-supervised node classification task, where a key point lies how to sufficiently leverage limited but valuable label information. Most of classical GNNs solely use known labels for computing loss at output. In recent years, several methods designed additionally utilize input. One part augment features via concatenating or adding them with one-hot encodings labels, while other optimize graph structure by assuming neighboring nodes tend same label. To bring into full play rich information this paper, we present label-enhanced learning framework GNNs, which first models each as virtual center intra-class and then jointly learns representations both labels. Our approach could not only smooth belonging class, also explicitly encode semantics process GNNs. Moreover, training selection technique is provided eliminate potential leakage issue guarantee model generalization ability. Finally, an adaptive self-training strategy proposed iteratively enlarge set more reliable pseudo distinguish importance pseudo-labeled during process. Experimental results on real-world synthetic datasets demonstrate our can consistently outperform state-of-the-arts, effectively nodes.