Table 4 Extended ablation study on UTD-MHAD dataset showing the impact of removing or isolating key components from the XTinyHAR model across multiple evaluation metrics.
Model variant | Test accuracy (%) | F1-score (%) | Precision (%) | Recall (%) | Cohen’s Kappa | FLOPs (M) |
|---|---|---|---|---|---|---|
Full model (KD + PE + AR) | 98.71 | 98.71 | 98.72 | 98.71 | 0.985 | 11.3 |
Without knowledge distillation (KD) | 96.84 | 96.80 | 96.87 | 96.84 | 0.962 | 11.3 |
Without positional embedding | 97.42 | 97.38 | 97.40 | 97.42 | 0.973 | 11.3 |
Without attention rollout | 98.11 | 98.08 | 98.12 | 98.11 | 0.978 | 11.3 |
Only knowledge distillation (KD only) | 96.41 | 96.40 | 96.45 | 96.41 | 0.960 | 11.3 |
Only positional embedding (PE only) | 95.78 | 95.73 | 95.70 | 95.78 | 0.954 | 11.3 |
Only attention rollout (AR only) | 95.01 | 94.94 | 94.99 | 95.01 | 0.944 | 11.3 |
