Fig. 2: The predictions of slavery prevalence (individuals enslaved as a % of the population) made by the best model using leave-one-out cross validation (LOOCV), compared to the “actual” prevalence as estimated using the Gallup World Poll (GWP) survey data.
From: Machine learning methods for “wicked” problems: exploring the complex drivers of modern slavery
The grey box plots illustrate the distribution of 10,000 bootstrapped LOOCV predictions (using the full pipeline NMF→DT) to help illustrate the uncertainty associated with our model’s predictions. The box shows the quartiles of the bootstrapped predictions while the whiskers extend to show the rest of the distribution, except for points that were determined to be “outliers” (using a function of the inter-quartile range), which are not plotted. The x-axis is ordered by the MAE.
