| Dyadic Model | Systemic Model | Ceiling Model | |
|---|---|---|---|
| node / endge encoder | 500 → 128 → 64 | 500 → 128 → 64 | 500 → 128 → 64 |
| edge classifier | 64 → 64 | 64 → 6 | 64 → 6 |
| learning rate | 0.001 | 0.001 | 0.001 |
| message passing steps | - | 2 | 2 |
| batch size | 512 | 512 | 512 |
| epochs | 30 | 30 | 30 |
| early stopping patience | 3 | 3 | 3 |
The grid search involved sweeping over the following hyperparameter space:
| Options | |
|---|---|
| node / endge encoder last layer | 32, 64, 128, 256, 512 |
| edge classifier last layer | 6, 12, 24, 48, 96 |
| learning rate | 0.01, 0.001, 0.0001 |
| message passing steps | 1,2,3 |