Divide and Denoise: Empowering Simple Models for Robust Semi-Supervised Node Classification against Label Noise, KDD 2024
DnD-NeT offers a new solution to tackle the two problems from both the model architecture and algorithm perspectives, reviving the utility of message passing and pseudo labels in the problem of semi-supervised node classification with noisy labels. Specifically, DnD-NeT involves a label-noise robust GNN equipped with a reliable graph pseudo labeling algorithm, which can attain both effectiveness and efficiency when solving the studied 5524 problem. Extensive experiments demonstrate its state-of-the-art performance in semi-supervised node classification with varying levels of label noise.
To run the code:
python main.pyRun on ogbn-arxiv:
python main_arix.py