Fig 1.
Maps showing the locations and relative weights given to presence points, randomly generated background points and target-group sampled background points.
Areas with no points are filled in black. In the plot, all point locations are jittered for pseudonymisation. Country boundaries source: Office for National Statistics licensed under the Open Government Licence v.3.0.
Fig 2.
Wilkinson dot plots and intervals showing the distribution of the probability assigned to points in the 2024 testing data by the ensemble model, split by class (whether the point was a true tick presence point or a background point).
Each bin represents 1% of the distribution, and each bin represents an equal number of observations (Kay 2023). The square point shows the median, and the thick black horizontal line shows the central two quartiles of the distribution. The dashed vertical line represents the 50% threshold.
Table 1.
Out-of-sample predictive performance metrics for the simple ensemble and base models.
Fig 3.
Modelled presence probabilities and actual tick reports from the 2024 testing data.
To prevent reidentification of exact locations, reports were aggregated in a 10km x 10km grid and are shown at the centroid of the grid square. Larger dots represent multiple reports in the same grid square. Country boundaries source: Office for National Statistics licensed under the Open Government Licence v.3.0.
Fig 4.
Permutation-based variable importance for simple ensemble predictions on the testing set.
The plot shows the ten most important variables on average across 25 iterations, along with confidence intervals. Loss is measured as impact on 1-ROC AUC.
Fig 5.
Partial dependence plots (PDPs) showing the average presence probability from the simple ensemble model, for the full range of the ten most important predictors in the testing data.
All other predictors are held at their mean in each plot.