Table 3 Comparison of ablation experiments.
From: Multi-scale spatial-temporal transformer for traffic flow prediction
Dataset | Models | Gate embedding | Two-stage spatial attention | Frequency dual-channel attention | MAE | RMSE | MAPE (%) |
|---|---|---|---|---|---|---|---|
PeMs03 | w/o gate embedding | \(\times\) | \(\checkmark\) | \(\checkmark\) | 14.48 | 25.73 | 14.24 |
w/o two-stage spatial attention | \(\checkmark\) | \(\times\) | \(\checkmark\) | 14.61 | 26.08 | 14.13 | |
w/o frequency dual-channel attention | \(\checkmark\) | \(\checkmark\) | \(\times\) | 14.56 | 25.82 | 14.16 | |
MSSTFormer | \(\checkmark\) | \(\checkmark\) | \(\checkmark\) | 14.35 | 25.01 | 14.05 | |
PeMs04 | w/o gate embedding | \(\times\) | \(\checkmark\) | \(\checkmark\) | 18.05 | 30.11 | 11.71 |
w/o two-stage spatial attention | \(\checkmark\) | \(\times\) | \(\checkmark\) | 18.18 | 30.46 | 11.59 | |
w/o frequency dual-channel attention | \(\checkmark\) | \(\checkmark\) | \(\times\) | 18.11 | 30.21 | 11.64 | |
MSSTFormer | \(\checkmark\) | \(\checkmark\) | \(\checkmark\) | 17.93 | 29.40 | 11.52 | |
PeMs07 | w/o gate embedding | \(\times\) | \(\checkmark\) | \(\checkmark\) | 18.92 | 33.18 | 8.11 |
w/o two-stage spatial attention | \(\checkmark\) | \(\times\) | \(\checkmark\) | 19.05 | 33.55 | 7.98 | |
w/o frequency dual-channel attention | \(\checkmark\) | \(\checkmark\) | \(\times\) | 18.96 | 33.27 | 8.02 | |
MSSTFormer | \(\checkmark\) | \(\checkmark\) | \(\checkmark\) | 18.79 | 32.48 | 7.89 | |
PeMs08 | w/o gate embedding | \(\times\) | \(\checkmark\) | \(\checkmark\) | 13.31 | 24.04 | 8.92 |
w/o two-stage spatial attention | \(\checkmark\) | \(\times\) | \(\checkmark\) | 13.42 | 24.33 | 8.84 | |
w/o frequency dual-channel attention | \(\checkmark\) | \(\checkmark\) | \(\times\) | 13.38 | 24.15 | 8.89 | |
MSSTFormer | \(\checkmark\) | \(\checkmark\) | \(\checkmark\) | 13.20 | 23.31 | 8.76 |
