Training Graph Convolutional Neural Network Against Label Noise

For node classification task, graph convolutional neural network (GCN) has achieved competitive performance on graph-structured data. Under semi-supervised setting, only a small portion of nodes are labeled for training. Many existing works have a perfect assumption that all the class labels used for training are completely accurate. However, noises are inevitably involved in the process of labeling, which can cause a degraded model performance. Yet few works focus on how to deal with noisy labels on graph data. Techniques against label noise on image domain can’t be applied to graph data directly. In this paper, we propose a framework, called super-nodes assisted label correction and dynamic graph adjustment based GCN (SuLD-GCN), which aims to reduce the negative impact of noise via label correction to obtain a higher-quality labels. We introduce the super-node to construct a new graph, which contributes to connecting nodes with the same class label more strongly. During iterations, we select nodes with high predicted confidence to correct their labels. Simultaneously, we adjust the graph structure dynamically. Experiments on public datasets demonstrate the effectiveness of our proposed method, yielding a significant improvement over state-of-art baselines.