A noise-resistant graph neural network by semi-supervised contrastive learning

Graph neural networks (GNNs) have been widely applied for representation learning on the graph data in real applications, but few of them are designed to conduct representation learning on the graph data with noisy labels. Its key challenge is that the feature embeddings of nodes with noisy labels (noisy nodes for short) are close to those of unlabeled nodes so that the classifier constructed by GNNs is influenced by noisy nodes. To address this issue, in this paper, we propose a noise-resistant graph neural network with semi-supervised contrastive learning to push noisy nodes far away from unlabeled nodes in the embedding space. To do this, we design a constraint of semi-supervised contrastive learning and put it into the objective function of GNNs. Specifically, the proposed constraint enlarges the distance between noisy nodes and unlabeled nodes by pushing noisy nodes far away from their unlabeled neighbors in the embedding space. As a result, the embeddings of unlabeled nodes are influenced by noisy label less. Moreover, we intuitively analyze the feasibility of our proposed constraint. Comprehensive experiments on real datasets further verify the effectiveness of our proposed method over previous SOTA methods in terms of classification tasks with different ratio levels of noisy labels.