HSG-12M
HSG-topology
Structure
1,812,325Graphs
1401Classes
4Node Feature Dim
6Edge Feature Dim
Coverage
1,401-class graph classification Classify each topology-deduplicated Hamiltonian spectral graph into one of 1,401 characteristic-polynomial classes. The variant is designed around unique connectivity patterns. Evaluated by Accuracy.
Dataset details
Task
Classify each topology-deduplicated Hamiltonian spectral graph into one of 1,401 characteristic-polynomial classes.
Data
HSG-topology retains one representative of every graph-isomorphism class within each polynomial class, leaving 1,812,325 spatial multigraphs with a largest-to- smallest class-size ratio of 660.2.
Features
The variant is designed around unique connectivity patterns. Spatial multi-edge information remains available in the HSG representation, but topology-only models can deliberately omit it for the combinatorial experiment.
Splits and evaluation
The official stratified random split is 80/10/10 and evaluation reports accuracy averaged over three seeds.
Quirks and pitfalls
This is the only deliberately imbalanced HSG static variant. Plain accuracy is dominated by large classes; balanced metrics are useful diagnostics but are not the published headline. Isomorphism filtering must be performed within each class.
Sources
SOTA Progression
new SOTA
matches SOTA at that time
other proposed result
other baseline result
Leaderboard
| # | Model | Accuracy ↑ | Loss | Macro F1 | Top-10 Acc. | Top-5 Accuracy | Paper |
|---|---|---|---|---|---|---|---|
| 1 | GraphSAGE | 0.6200
±
0.0030 Loss
1.0190
±
0.0020 Macro F1
0.6220
±
0.0010 Top-10 Acc.
0.9580
±
0.0010 Top-5 Accuracy
0.8980
±
0.0020 | 1.0190
±
0.0020 | 0.6220
±
0.0010 | 0.9580
±
0.0010 | 0.8980
±
0.0020 | |
| 2 | CGCNN | 0.5660
±
0.0160 Loss
1.1910
±
0.0650 Macro F1
0.5630
±
0.0180 Top-10 Acc.
0.9400
±
0.0050 Top-5 Accuracy
0.8700
±
0.0080 | 1.1910
±
0.0650 | 0.5630
±
0.0180 | 0.9400
±
0.0050 | 0.8700
±
0.0080 | |
| 3 | GINE | 0.5330
±
0.0170 Loss
1.3160
±
0.0570 Macro F1
0.5310
±
0.0110 Top-10 Acc.
0.9270
±
0.0080 Top-5 Accuracy
0.8480
±
0.0130 | 1.3160
±
0.0570 | 0.5310
±
0.0110 | 0.9270
±
0.0080 | 0.8480
±
0.0130 | |
| 4 | GAT | 0.4340
±
0.0150 Loss
1.8250
±
0.0530 Macro F1
0.4310
±
0.0110 Top-10 Acc.
0.8550
±
0.0100 Top-5 Accuracy
0.7520
±
0.0140 | 1.8250
±
0.0530 | 0.4310
±
0.0110 | 0.8550
±
0.0100 | 0.7520
±
0.0140 | |
| 5 | GATv2 | 0.4010
±
0.0030 Loss
1.9680
±
0.0250 Macro F1
0.3970
±
0.0010 Top-10 Acc.
0.8330
±
0.0050 Top-5 Accuracy
0.7230
±
0.0050 | 1.9680
±
0.0250 | 0.3970
±
0.0010 | 0.8330
±
0.0050 | 0.7230
±
0.0050 | |
| 6 | GCN | 0.3970
±
0.0090 Loss
2.0620
±
0.0620 Macro F1
0.3920
±
0.0110 Top-10 Acc.
0.8250
±
0.0060 Top-5 Accuracy
0.7150
±
0.0070 | 2.0620
±
0.0620 | 0.3920
±
0.0110 | 0.8250
±
0.0060 | 0.7150
±
0.0070 | |
| 7 | MF | 0.3480
±
0.0120 Loss
2.2220
±
0.0510 Macro F1
0.3430
±
0.0120 Top-10 Acc.
0.7930
±
0.0100 Top-5 Accuracy
0.6730
±
0.0110 | 2.2220
±
0.0510 | 0.3430
±
0.0120 | 0.7930
±
0.0100 | 0.6730
±
0.0110 | |
| 8 | GIN | 0.0950
±
0.0590 Loss
4.1790
±
0.6120 Macro F1
0.0820
±
0.0600 Top-10 Acc.
0.3900
±
0.1480 Top-5 Accuracy
0.2760
±
0.1270 | 4.1790
±
0.6120 | 0.0820
±
0.0600 | 0.3900
±
0.1480 | 0.2760
±
0.1270 |
SOTA performance
set a new SOTA at the time
matched the SOTA at the time