RLGNN: Robust Graph Neural Networks with Noisy Label Learning

Haodong Qian, De Li, Qiyu Li, Jinyan Wang

Abstract: Graph Neural Networks (GNNs) have become a powerful tool for processing graph-structured data. However, real-world data often faces the problem of label noise and sparsity, which will greatly affect the application of graph neural networks in the real world. Hence, we introduce Robust Graph Neural Networks with Noisy Label Learning (RLGNN) to tackle the challenges posed by label noise and sparsity in graph-structured data. Its key components include unsupervised contrastive learning for deriving robust node representations, a two-stage clean sample selection mechanism, and a semi-supervised loss function to harness information from unlabeled nodes. RLGNN ingeniously merges contrastive learning with the prediction space, refining the label selection process and boosting the model’s robustness and generalizability. Comprehensive experiments indicate that RLGNN significantly surpasses existing models in managing label noise in semi-supervised node classification scenarios.