Extended Data Fig. 6: Hyper-parameter selection for the XGBoost model.
From: Co-limitation towards lower latitudes shapes global forest diversity gradients
Using 20 bootstrapping iterations on random training sets consisting of 90% of sample for each continent, we calculated the sensitivity of the root-mean-squared error (RMSE) of the testing sets (consisting of the remaining 10% of sample) to (left) the maximum number of boosting iterations (that is number of rounds), and (right) the maximum depth of a tree for the XGBoost model. As RMSE converged at 50 rounds and 20 depth, we selected them as optimal hyper-parameters for the XGBoost model. The box plot represents the 25th and 75th percentiles (bounds of box), median (centre line), and the maximum and minimum (upper and lower whiskers) of the RMSE values for each level of tree depth.
