An Open Access Journal
Table 4 Comparison results for different model settings and existing approach
From: Detection of anomalies in cycling behavior with convolutional neural network and deep learning
Validation | Threshold | Performance metrics | ||
|---|---|---|---|---|
Recall | Precision | F2 | ||
B1) Threshold = 0.15 | 0.15 | 0.80 | 0.54 | 0.73 |
B2) Threshold = 0.21 | 0.21 | 0.46 | 0.68 | 0.49 |
C1) TS_window_size = 64 | 0.18 | 0.68 | 0.59 | 0.66 |
C2) TS_window_size = 80 | 0.18 | 0.56 | 0.59 | 0.57 |
D) No SGF Filter + speed and Heading parameters | 0.18 | 0.95 | 0.36 | 0.72 |
E) SGF + Only Speed parameters | 0.18 | 0.61 | 0.57 | 0.60 |
F) SGF + Only Heading parameters | 0.18 | 0.54 | 0.67 | 0.56 |
A1) Training each user (80/20) TS_window_size = 40, SGH + Speed parameters + Heading parameters | 0.21 | 0.83 | 0.60 | 0.77 |
A2) Weighted Average of A2 | 0.75 | 0.62 | 0.72 | |
PCA—training each user (80/20), TS_window_size = 40 | 0.17 | 0.66 | 0.34 | 0.45 |
PCA—ID-4 training (100%), TS_window_size = 40, SGH + Speed parameters + Heading parameters | 0.17 | 0.49 | 0.44 | 0.47 |
Breaking Threshold—ID-4 training (100%), only Speed parameters | variable | < 0.30 | < 0.30 | < 0.30 |