Learning on Graphs under Label Noise
This addresses a practical issue in graph-based applications like social analysis where labels are often inaccurate, offering a robust solution for noisy label scenarios.
The paper tackles the problem of node classification on graphs when labels are noisy, developing a Consistent Graph Neural Network (CGNN) that uses graph contrastive learning for robustness and a homophily-based technique to detect and purify noisy labels, achieving superior performance on three benchmark datasets.
Node classification on graphs is a significant task with a wide range of applications, including social analysis and anomaly detection. Even though graph neural networks (GNNs) have produced promising results on this task, current techniques often presume that label information of nodes is accurate, which may not be the case in real-world applications. To tackle this issue, we investigate the problem of learning on graphs with label noise and develop a novel approach dubbed Consistent Graph Neural Network (CGNN) to solve it. Specifically, we employ graph contrastive learning as a regularization term, which promotes two views of augmented nodes to have consistent representations. Since this regularization term cannot utilize label information, it can enhance the robustness of node representations to label noise. Moreover, to detect noisy labels on the graph, we present a sample selection technique based on the homophily assumption, which identifies noisy nodes by measuring the consistency between the labels with their neighbors. Finally, we purify these confident noisy labels to permit efficient semantic graph learning. Extensive experiments on three well-known benchmark datasets demonstrate the superiority of our CGNN over competing approaches.