Fortifying Robustness in Graph Neural Networks: A Loss Correction Approach to Mitigate Label Noise.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Noisy labels in training data can significantly degrade neural network performance.
• Graph Neural Networks (GNNs) are susceptible to label noise due to their information propagation mechanisms.
• Existing robust GNNs often fail to adequately address label noise during training or edge data corruption.

Purpose of the Study:

• To develop a novel robust GNN framework, NomiGNN, to enhance resilience against label noise in node classification tasks.
• To mitigate the adverse effects of noisy labels and erroneous edge aggregation in GNNs.
• To improve the accuracy and reliability of GNNs in real-world graph datasets with corrupted labels.

Main Methods:

• Introduced noise-mitigating GNN (NomiGNN) framework with noise distribution estimation and refined loss optimization.
• Incorporated edge labels for a novel prediction task to learn sample relationships via same-label probabilities.
• Utilized pseudoedge labeling and iterative learning to address label shortages and estimation inaccuracies.

Main Results:

• NomiGNN demonstrated superior resilience against noisy label corruption compared to eight benchmark GNN models.
• Experimental evaluations on five real-world graphs validated the framework's effectiveness.
• The proposed methods successfully mitigated mis-aggregation from noisy edges and bolstered node classification accuracy.

Conclusions:

• NomiGNN offers a robust solution for training GNNs with noisy labels, significantly improving node classification performance.
• The framework effectively addresses limitations of existing robust GNNs by considering both label noise and edge data integrity.
• NomiGNN provides a promising direction for developing more reliable GNNs in practical applications with imperfect data.