Table 4:
Detailed accuracy for ablation experiments using LOMO evaluation. All tests are on FL data. Last column repeats LOMO results from Table 1 to facilitate comparison. RF is random forest, SSH is our semi-supervised hierarchical classifier.
| Train lab, RF |
Train lab+FL, RF |
Train lab, SSH |
Train lab+FL, SSH |
|||||
|---|---|---|---|---|---|---|---|---|
| % Total | Acc | % Total | Acc | % Total | Acc | % Total | Acc | |
| Level 1 | 40.8 | 68.2 | 45.4 | 70.3 | 29.9 | 82.2 | 35.8 | 86.4 |
| Level 2 | 36.7 | 61.8 | 30.5 | 71.2 | 35.1 | 80.1 | 30.2 | 85.5 |
| Level 3 | 15.5 | 50.3 | 17.5 | 60.9 | 26.2 | 69.9 | 28.7 | 77.8 |
| Level 4 | 7.0 | 48.9 | 6.7 | 61.1 | 8.9 | 70.4 | 5.4 | 74.9 |
|
| ||||||||
| Overall (intakes) | 61.7 | 68.4 | 77.3 | 83.1 | ||||
|
| ||||||||
| Meal-level (unweighted) | 59.3 | 68.9 | 71.9 | 76.5 | ||||
| Meal-level (weighted) | 59.9 | 68.2 | 76.7 | 80.1 | ||||