ZINC
Description
ZINC
Dataset Overview
Lifting Methods
Structural-based Liftings
- Cellular: Cycle-based lifting
- Simplicial: Clique complex lifting
- Hypergraph: k-hop lifting
Feature Lifting
- Projected sum
Model Performance
| Model | Accuracy (MAE) | Std Dev (±) |
|---|---|---|
| GCN | 0.62 | 0.01 |
| GIN | 0.57 | 0.04 |
| GAT | 0.61 | 0.01 |
| AST | 0.59 | 0.02 |
| EDGNN | 0.51 | 0.01 |
| UniGNN2 | 0.60 | 0.01 |
| CWN | 0.34 | 0.01 |
| CCCN | 0.34 | 0.02 |
| SCCNN | 0.36 | 0.02 |
| SCN | 0.53 | 0.04 |
Key Insights
- CWN and CCCN achieve the best performance with 0.34 accuracy
- GCN achieves the highest accuracy among traditional GNNs (0.62)
- Most models perform in the 0.5–0.6 range, with CWN and CCCN significantly outperforming others
- GIN shows the highest variability (±0.04), while GCN and EDGNN are the most stable (±0.01)
Reproducibility
python -m topobench model=cell/cwn dataset=graph/ZINC optimizer.parameters.lr=0.001 model.feature_encoder.out_channels=64 model.backbone.n_layers=2 model.readout.readout_name=PropagateSignalDown model.feature_encoder.proj_dropout=0.25 dataset.dataloader_params.batch_size=128 transforms.graph2cell_lifting.max_cell_length=10 callbacks.early_stopping.min_delta=0.005 transforms.one_hot_node_degree_features.degrees_fields=x seed=42,3,5,23,150 trainer.max_epochs=500 trainer.min_epochs=50 trainer.check_val_every_n_epoch=5 callbacks.early_stopping.patience=10 logger.wandb.project=TopoBench --multirun