Table 1.
Performance of different methods on BDG2 and CBTs in terms of accuracy, memory, training time, and communication overhead per round
| Method | BDG2 | CBTs | Memory (KB) | Training Time (s) | Communication (KB) | ||||
|---|---|---|---|---|---|---|---|---|---|
| RMSE | MAPE | MAE | RMSE | MAPE | MAE | ||||
| Cen | 22.82 | 8.41 | 8.20 | 0.4780 | 26.98 | 0.2727 | 2578.42 (1.0 ×) | 586.42 (2.05 ×) | – |
| Local-S | 23.56 | 8.13 | 8.12 | 0.4698 | 28.19 | 0.2726 | 152.53 (16.9 ×) | 41.19 (29.14 ×) | – |
| FedAvg-S | 22.79 | 7.67 | 7.98 | 0.4681 | 27.04 | 0.2699 | 152.53 (16.9 ×) | 46.49 (28.80 ×) | 5.50 (27.18 ×) |
| FedProx-S | 22.58 | 7.71 | 7.91 | 0.4668 | 27.17 | 0.2678 | 152.53 (16.9 ×) | 46.49 (28.80 ×) | 5.50 (27.18 ×) |
| Local-M | 22.84 | 7.75 | 8.03 | 0.4645 | 26.67 | 0.2682 | 2578.42 (1.0 ×) | 1187.15 (1.01 ×) | – |
| FedAvg-M | 22.25 | 6.96 | 7.48 | 0.4614 | 25.81 | 0.2637 | 2578.42 (1.0 ×) | 1200.44 (1.0 ×) | 149.50 (1.0 ×) |
| FedProx-M | 22.21 | 7.16 | 7.62 | 0.4622 | 25.76 | 0.2639 | 2578.42 (1.0 ×) | 1200.44 (1.0 ×) | 149.50 (1.0 ×) |
| Split | 23.27 | 7.68 | 7.96 | 0.4679 | 26.83 | 0.2687 | 103.75 (24.8 ×) | 96.00 (12.50 ×) | 91.25 (1.63 ×) |
| SFLV1 | 22.34 | 7.25 | 7.68 | 0.4647 | 26.03 | 0.2664 | 103.75 (24.8 ×) | 96.49 (12.44 ×) | 96.75 (1.54 ×) |
| SFLV2 | 22.76 | 7.53 | 7.89 | 0.4674 | 26.49 | 0.2683 | 103.75 (24.8 ×) | 96.49 (12.44 ×) | 96.75 (1.54 ×) |
| Proposed | 22.17 | 6.98 | 7.44 | 0.4630 | 25.74 | 0.2636 | 115.07 (22.4 ×) | 62.41 (19.23 ×) | 74.43 (2.01 ×) |
1The best-performing and second-best-performing methods are bolded and underlined, respectively.
2-S and -M indicates that the model has single and multiple hidden layers, respectively.