Coverage approaches observed retraining agreement
92%
of the observed retraining agreement (ρ = 0.57 / 0.62), using training-free source ranking
9.0
MAE in PCK for an unseen source on a known target (LOSO, r = 0.89)
| Held-out split | Generalizes to | r | MAE (PCK) |
|---|---|---|---|
| LOSO — hold source | unseen source, known target | 0.89 | 9.0 |
| LOTO — hold target | unseen target | 0.73 | 16.6 |
| LOSTO — hold both | fully novel pair | 0.68 | 19.4 |
More: the ceiling plot and the absolute-prediction scatter2 figures
