Table 5
Hyper-parameter space used for tuning the XGBoost model.
| Hyper-parameter | Min. value | Max. value | 2 det. | 3 det. | 4 det. | 5 det. | 6 det. | 7 to 100 det. |
|---|---|---|---|---|---|---|---|---|
| Column sample by level | 0.01 | 1.0 | 0.01 | 1.00 | 1.00 | 1.00 | 1.00 | 0.00 |
| Column sample by tree | 0.1 | 1.0 | 0.74 | 0.75 | 0.69 | 0.92 | 0.75 | 0.74 |
| Gamma | 0.0 | 40.0 | 0.14 | 1.65 | 3.81 | 2.49 | 1.65 | 0.13 |
| Learning rate | 0.0005 | 0.5 | 0.38 | 0.17 | 0.10 | 0.32 | 0.17 | 0.38 |
| Maximum depth | 2 | 12 | 12 | 9 | 3 | 11 | 9 | 12 |
| Minimum child node weight | 0.0001 | 0.5 | 0.32 | 0.04 | 0.42 | 0.21 | 0.04 | 0.32 |
| Subsample | 0.01 | 1.0 | 0.79 | 0.48 | 0.60 | 0.36 | 0.48 | 0.79 |
Notes. The tuning process was done via a randomized search, and final optimal hyper-parameters obtained per model by number of detections. 200 points were randomly selected from the joint distribution of uniform probabilities within each hyper-parameter range. A five-fold cross validation was then performed for each of them and the mean performance was calculated. The point with the highest mean performance was selected as the optimum combination of hyper-parameter values.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.