Table 1.
The accuracy of transfer learning and direct validation
| Train Net. | Fitness (N = 500) | Fitness (N = 1000) | |||
|---|---|---|---|---|---|
| Test Net. | xT | xD | xT | xD | |
| BA (N = 500) | 0.853 ± 0.004 | 0.694 ± 0.03 | 0.859 ± 0.001 | 0.697 ± 0.012 | |
| BA (N = 1000) | 0.832 ± 0.003 | 0.685 ± 0.020 | 0.839 ± 0.001 | 0.679 ± 0.013 | |
| PSO (N = 500) | 0.830 ± 0.007 | 0.682 ± 0.025 | 0.845 ± 0.002 | 0.653 ± 0.002 | |
| PSO (N = 1000) | 0.836 ± 0.007 | 0.701 ± 0.019 | 0.848 ± 0.001 | 0.664 ± 0.019 | |
"Train Net.” and “Test Net.” refer to the networks used to train and test the ensemble models, respectively. The pairwise accuracy of the ensemble model evaluated on “Test Net.” under transfer learning is denoted as xT and that under direct validation is denoted as xD. The network models used are the Barabási–Albert (BA) model15, the popularity-similarity-optimization (PSO) model38,53, and the fitness model39 (detailed information can be found in Supplementary Section 8). The value in the table represents the average accuracy and its standard deviation from 10 simulations. Better results are highlighted in boldface. Swapping the training and testing network models yields consistent results.