The leaderboard includes the best performing GNN models on each datasets, in order, with their scores and the number of trainable parameters.
Models with configs having 500k trainable parameters
| Rank | Model | #Params | Test Acc ± s.d. | Links |
|---|---|---|---|---|
| 1 | GatedGCN-PE | 505421 | 86.363 ± 0.127 | Paper |
| 2 | RingGNN | 504766 | 86.244 ± 0.025 | Paper |
| 3 | MoNet | 511487 | 85.582 ± 0.038 | Paper |
| 4 | GatedGCN | 502223 | 85.568 ± 0.088 | Paper |
| 5 | GIN | 508574 | 85.387 ± 0.136 | Paper |
| 6 | 3WLGNN | 502872 | 85.341 ± 0.207 | Paper |
| 7 | GAT | 526990 | 78.271 ± 0.186 | Paper |
| 8 | GCN | 500823 | 71.892 ± 0.334 | Paper |
| 9 | GraphSage | 502842 | 50.492 ± 0.001 | Paper |
Models with configs having 500k trainable parameters
| Rank | Model | #Params | Test Acc ± s.d. | Links |
|---|---|---|---|---|
| 1 | GatedGCN-PE | 503473 | 74.088 ± 0.344 | Paper |
| 2 | GatedGCN | 502615 | 73.840 ± 0.326 | Paper |
| 3 | GAT | 527874 | 70.587 ± 0.447 | Paper |
| 4 | GCN | 501687 | 68.498 ± 0.976 | Paper |
| 5 | MoNet | 511999 | 66.407 ± 0.540 | Paper |
| 6 | GIN | 517570 | 64.716 ± 1.553 | Paper |
| 7 | GraphSage | 503350 | 63.844 ± 0.110 | Paper |
| 8 | 3WLGNN | 507252 | 55.489 ± 7.863 | Paper |
| 9 | RingGNN | 524202 | 22.340 ± 0.000 | Paper |
Models with configs having 500k trainable parameters
| Rank | Model | #Params | Test MAE ± s.d. | Links |
|---|---|---|---|---|
| 1 | PNA | 387155 | 0.142 ± 0.010 | Paper, Code |
| 2 | MPNN (sum) | 480805 | 0.145 ± 0.007 | Paper, Code |
| 3 | GatedGCN-PE | 505011 | 0.214 ± 0.006 | Paper |
| 4 | MPNN (max) | 480805 | 0.252 ± 0.009 | Paper, Code |
| 5 | GatedGCN-E | 504309 | 0.282 ± 0.015 | Paper |
| 6 | MoNet | 504013 | 0.292 ± 0.006 | Paper |
| 7 | 3WLGNN-E | 507603 | 0.303 ± 0.068 | Paper |
| 8 | RingGNN-E | 527283 | 0.353 ± 0.019 | Paper |
| 9 | GCN | 505079 | 0.367 ± 0.011 | Paper |
| 10 | GAT | 531345 | 0.384 ± 0.007 | Paper |
| 11 | GraphSage | 505341 | 0.398 ± 0.002 | Paper |
| 12 | GIN | 509549 | 0.526 ± 0.051 | Paper |
Models with configs having 100k trainable parameters
| Rank | Model | #Params | Test Acc ± s.d. | Links |
|---|---|---|---|---|
| 1 | PNA | 119812 | 97.940 ± 0.120 | Paper, Code |
| 2 | MPNN (max) | 109057 | 97.690 ± 0.220 | Paper, Code |
| 3 | GatedGCN | 104217 | 97.340 ± 0.143 | Paper |
| 4 | GraphSage | 104337 | 97.312 ± 0.097 | Paper |
| 5 | MPNN (sum) | 109057 | 96.900 ± 0.150 | Paper, Code |
| 6 | GIN | 105434 | 96.485 ± 0.252 | Paper |
| 7 | GAT | 110400 | 95.535 ± 0.205 | Paper |
| 8 | 3WLGNN | 108024 | 95.075 ± 0.961 | Paper |
| 9 | MoNet | 104049 | 90.805 ± 0.032 | Paper |
| 10 | GCN | 101365 | 90.705 ± 0.218 | Paper |
| 11 | RingGNN | 105398 | 11.350 ± 0.000 | Paper |
Models with configs having 500k trainable parameters for 3WLGNN and RingGNN
| Rank | Model | #Params | Test Acc ± s.d. | Links |
|---|---|---|---|---|
| 1 | 3WLGNN | 501690 | 95.002 ± 0.419 | Paper |
| 2 | RingGNN | 505182 | 91.860 ± 0.449 | Paper |
Models with configs having 100k trainable parameters
| Rank | Model | #Params | Test Acc ± s.d. | Links |
|---|---|---|---|---|
| 1 | MPNN (max) | 109277 | 70.860 ± 0.270 | Paper, Code |
| 2 | PNA | 113472 | 70.350 ± 0.630 | Paper, Code |
| 3 | GatedGCN | 104357 | 67.312 ± 0.311 | Paper |
| 4 | GraphSage | 104517 | 65.767 ± 0.308 | Paper |
| 5 | MPNN (sum) | 109277 | 65.610 ± 0.300 | Paper, Code |
| 6 | GAT | 110704 | 64.223 ± 0.455 | Paper |
| 7 | 3WLGNN | 108516 | 59.175 ± 1.593 | Paper |
| 8 | GCN | 101657 | 55.710 ± 0.381 | Paper |
| 9 | GIN | 105654 | 55.255 ± 1.527 | Paper |
| 10 | MoNet | 104229 | 54.655 ± 0.518 | Paper |
| 11 | RingGNN | 105165 | 19.300 ± 16.108 | Paper |
Models with configs having 500k trainable parameters for 3WLGNN and RingGNN
| Rank | Model | #Params | Test Acc ± s.d. | Links |
|---|---|---|---|---|
| 1 | 3WLGNN | 502770 | 58.043 ± 2.512 | Paper |
| 2 | RingGNN | 504949 | 39.165 ± 17.114 | Paper |
Models with configs having 100k trainable parameters
| Rank | Model | #Params | Test F1 ± s.d. | Links |
|---|---|---|---|---|
| 1 | GatedGCN-E | 97858 | 0.808 ± 0.003 | Paper |
| 2 | GatedGCN | 97858 | 0.791 ± 0.003 | Paper |
| 3 | 3WLGNN-E | 106366 | 0.694 ± 0.073 | Paper |
| 4 | k-NN baseline | NA(k=2) | 0.693 ± 0.000 | Paper |
| 5 | GAT | 96182 | 0.671 ± 0.002 | Paper |
| 6 | GraphSage | 99263 | 0.665 ± 0.003 | Paper |
| 7 | GIN | 99002 | 0.656 ± 0.003 | Paper |
| 8 | RingGNN-E | 106862 | 0.643 ± 0.024 | Paper |
| 9 | MoNet | 99007 | 0.641 ± 0.002 | Paper |
| 10 | GCN | 95702 | 0.630 ± 0.001 | Paper |
Models with configs having 500k trainable parameters
| Rank | Model | #Params | Test F1 ± s.d. | Links |
|---|---|---|---|---|
| 1 | GatedGCN-E | 500770 | 0.838 ± 0.002 | Paper |
| 2 | RingGNN-E | 507938 | 0.704 ± 0.003 | Paper |
| 3 | k-NN baseline | NA(k=2) | 0.693 ± 0.000 | Paper |
| 4 | 3WLGNN-E | 506681 | 0.288 ± 0.311 | Paper |
Models with configs having 40k trainable parameters
| Rank | Model | #Params | Test Hits@50 ± s.d. | Links |
|---|---|---|---|---|
| 1 | GatedGCN | 40965 | 52.816 ± 1.303 | Paper |
| 2 | GatedGCN-PE | 42769 | 52.018 ± 1.178 | Paper |
| 3 | GraphSage | 39856 | 51.618 ± 0.690 | Paper |
| 4 | GAT | 42637 | 51.501 ± 0.962 | Paper |
| 5 | GCN | 40479 | 50.422 ± 1.131 | Paper |
| 6 | GatedGCN-E | 40965 | 49.212 ± 1.560 | Paper |
| 7 | MatrixFact baseline | - | 44.206 ± 0.452 | Paper |
| 8 | GIN | 39544 | 41.730 ± 2.284 | Paper |
| 9 | MoNet | 39751 | 36.144 ± 2.191 | Paper |
Note for OGBL-COLLAB
- 40k params is the highest we could fit the single OGBL-COLLAB graph on GPU for fair comparisons.
- RingGNN and 3WLGNN rely on dense tensors which leads to OOM on both GPU and CPU memory.