Graph Neural Networks with Coarse- and Fine-Grained Division for mitigating label noise and sparsity

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Graph Neural Networks (GNNs) excel in semi-supervised learning but struggle with noisy and sparse labels.
• Real-world graph data often suffers from label imperfections, degrading GNN performance.
• Robust GNNs for node classification under label noise are crucial.

Purpose of the Study:

• Propose GNN-CFGD, a novel GNN architecture to mitigate label sparsity and noise.
• Enhance the robustness of semi-supervised node classification on graphs.
• Improve GNN performance in practical, imperfect labeling scenarios.

Main Methods:

• Developed GNN-CFGD utilizing coarse- and fine-grained label division and graph reconstruction.
• Employed a Gaussian Mixture Model (GMM) with memory effect to identify clean labels.
• Introduced a clean label-oriented link to connect unlabeled nodes to clean ones.
• Fine-grained noisy and unlabeled nodes based on confidence for refined supervision.

Main Results:

• Demonstrated that linking unlabeled nodes to clean labels is more effective against noise.
• GNN-CFGD effectively reduces the impact of noisy labels through its division strategy.
• Experiments show superior effectiveness and robustness of GNN-CFGD across various datasets.

Conclusions:

• GNN-CFGD offers a robust solution for semi-supervised node classification with noisy and sparse labels.
• The proposed coarse- and fine-grained division strategy significantly improves GNN performance.
• This work addresses a critical challenge in applying GNNs to real-world graph data.