We sincerely appreciate all reviewers for their valuable comments. Based on the comments, we have conducted additional experiments and the results are reported as follows.
Additional results on replacing our label noise-robust GNN with non-decoupled GCN as the backbone are as follows:
| Dataset | Methods | Clean Labels | Uniform Noise 20% | Uniform Noise 40% | Uniform Noise 60% | Pair Noise 20% | Pair Noise 30% | Pair Noise 40% | Pair Noise 50% |
|---|---|---|---|---|---|---|---|---|---|
| Cora-ML | GCN+NPA+PL | 84.96±1.12 | 77.11±3.56 | 74.94±2.95 | 52.42±5.32 | 82.69±1.62 | 76.43±2.60 | 68.05±2.83 | 43.99±3.51 |
| DnD-NeT | 85.58±0.96 | 83.88±1.87 | 79.14±2.54 | 63.56±6.15 | 84.72±1.29 | 80.47±2.03 | 75.18±3.95 | 49.79±7.99 | |
| CiteSeer | GCN+NPA+PL | 73.24±1.23 | 69.31±2.29 | 61.69±6.90 | 42.74±6.14 | 67.56±3.26 | 63.93±4.54 | 61.64±4.68 | 37.39±5.54 |
| DnD-NeT | 75.57±1.14 | 74.30±1.20 | 70.85±2.17 | 58.35±5.37 | 73.98±1.22 | 71.36±2.31 | 70.26±2.82 | 46.33±3.06 | |
| PubMed | GCN+NPA+PL | 82.74±0.47 | 81.55±1.11 | 72.16±8.59 | 43.62±2.67 | 79.94±1.48 | 76.51±2.81 | 70.58±3.03 | 60.17±2.80 |
| DnD-NeT | 84.02±0.44 | 84.26±0.45 | 80.73±1.34 | 65.63±5.12 | 82.33±0.56 | 79.29±1.13 | 78.45±2.39 | 60.74±3.74 | |
| Coauthor CS | GCN+NPA+PL | 92.54±0.16 | 91.62±0.23 | 90.03±0.51 | 89.29±0.74 | 90.91±0.41 | 88.25±0.93 | 79.34±3.1 | 49.06±10.8 |
| DnD-NeT | 92.71±0.25 | 92.14±0.27 | 90.94±0.52 | 90.00±0.63 | 91.99±0.47 | 88.44±1.13 | 80.97±1.32 | 65.38±4.97 |
The results demonstrate that conventional non-decoupled GNN with NPA and PL performs behind DnD-NET, which shows the importance of our decoupled architecture.
The experiment results of all methods on five datasets with 50% pair noise are reported as follows:
| Dataset | Methods | Pair Noise 50% | Dataset | Methods | Pair Noise 50% |
|---|---|---|---|---|---|
| Cora-ML | GCN | 44.43±3.44 | CiteSeer | GCN | 37.94±4.03 |
| SGC | 48.80±5.32 | SGC | 39.66±3.79 | ||
| GRAND | 41.20±2.95 | GRAND | 30.19±7.78 | ||
| PTA | 42.37±5.15 | PTA | 44.37±6.24 | ||
| Co-teaching | 47.53±6.47 | Co-teaching | 41.11±5.29 | ||
| T-Revision | 47.33±4.30 | T-Revision | 41.42±4.79 | ||
| NRGNN | 48.23±4.23 | NRGNN | 42.15±3.62 | ||
| RTGNN | 43.66±4.07 | RTGNN | 37.30±4.72 | ||
| Pi-GNN | 45.03±2.71 | Pi-GNN | 40.02±4.55 | ||
| ERASE | 43.93±3.35 | ERASE | 45.89±7.36 | ||
| DnD-NeT | 49.79±7.99 | DnD-NeT | 46.33±3.06 |
| Dataset | Methods | Pair Noise 50% | Dataset | Methods | Pair Noise 50% |
|---|---|---|---|---|---|
| PubMed | GCN | 53.85±3.14 | Coauthor CS | GCN | 51.08±3.74 |
| SGC | 59.84±2.68 | SGC | 61.73±5.39 | ||
| GRAND | 60.26±3.56 | GRAND | 60.26±3.56 | ||
| PTA | 53.90±3.38 | PTA | 64.14±6.04 | ||
| Co-teaching | 55.02±8.14 | Co-teaching | 59.61±6.36 | ||
| T-Revision | 58.39±3.73 | T-Revision | 60.92±6.76 | ||
| NRGNN | 55.19±3.69 | NRGNN | 60.76±5.82 | ||
| RTGNN | 53.96±6.19 | RTGNN | 43.39±2.86 | ||
| Pi-GNN | 54.88±4.29 | Pi-GNN | 53.49±6.03 | ||
| ERASE | 57.33±3.83 | ERASE | 46.09±3.12 | ||
| DnD-NeT | 60.74±3.74 | DnD-NeT | 65.38±4.97 |
| Dataset | Methods | Pair Noise 50% |
|---|---|---|
| ogbn-arxiv | GCN | 38.96±2.4 |
| Co-teaching | 34.42±5.8 | |
| T-Revision | 32.18±5.8 | |
| Pi-GNN | 39.38±3.5 | |
| ERASE | 37.01±7.6 | |
| DnD-NeT | 40.34±1.2 |
As shown in the results, our approach can still achieve noticeable improvements over baselines on high label noise ratio on different datasets.
The comparison results between ERASE and our approach under the setting of ERASE are as follows:
| Cora | Uniform Noise 10% | Uniform Noise 20% | Uniform Noise 30% | Uniform Noise 40% | Uniform Noise 50% |
|---|---|---|---|---|---|
| ERASE | 81.58±0.80 | 80.37±0.77 | 79.61±0.95 | 78.13±1.07 | 78.01±1.05 |
| DnD-NeT | 82.40±0.43 | 82.10±0.39 | 79.70±1.41 | 78.72.2±1.57 | 78.23±1.38 |
| Cora | Pair Noise 10% | Pair Noise 20% | Pair Noise 30% | Pair Noise 40% | Pair Noise 50% |
|---|---|---|---|---|---|
| ERASE | 81.43±0.90 | 80.46±1.00 | 79.52±1.13 | 75.36±2.32 | 48.00±2.52 |
| DnD-NeT | 82.80±1.73 | 81.95±0.71 | 80.15±0.36 | 77.15±2.51 | 53.43±4.81 |
| PubMed | Uniform Noise 10% | Uniform Noise 20% | Uniform Noise 30% | Uniform Noise 40% | Uniform Noise 50% |
|---|---|---|---|---|---|
| ERASE | 78.16±0.88 | 77.86±1.07 | 75.60±0.85 | 72.72±0.79 | 70.72±1.08 |
| DnD-NeT | 79.43±0.89 | 78.70±1.99 | 76.11±0.95 | 74.30±1.26 | 71.45±1.17 |
| PubMed | Pair Noise 10% | Pair Noise 20% | Pair Noise 30% | Pair Noise 40% | Pair Noise 50% |
|---|---|---|---|---|---|
| ERASE | 78.01±1.10 | 76.70±1.07 | 76.71±1.55 | 73.00±2.62 | 61.46±2.90 |
| DnD-NeT | 79.58+1.87 | 78.73±0.75 | 77.13±1.03 | 72.60±2.69 | 67.48±2.02 |
Based on the results, we can observe that our model still outperforms ERASE under this setting.
Divide and Denoise: Empowering Simple Models for Robust Semi-Supervised Node Classification against Label Noise
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 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
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent