Correcting for the effects of class imbalance improves the performance of machine-learning based species distribution models

Numerous methods have been developed to combat the unwanted effects of imbalanced training data on the performance of machine-learning based predictive models. These methods attempt to balance model sensitivity and specificity. However, the effects of specific imbalance-correction methods on the performance of different machine-learning algorithms are not well understood for ecological data. In this study, we used four machine-learning algorithms (random forest, artificial neural network, gradient boosting, support vector machine) and five imbalance-correction methods (base algorithm = no correction, cutoff, up-sampling, down-sampling, weighting) to produce species distribution models for 15 freshwater macroinvertebrate genera that varied from 2.5 to 29.0% in prevalence. All imbalance-correction methods substantially improved average model performance (true skill statistic) over the base machine-learning algorithms, except when up-sampling was applied to random forest models. Choice of machine-learning algorithm had little effect on model performance, although gradient boosting performed better than other algorithms on the most imbalanced datasets. Our results suggest that the performance of species distribution models built with presence/absence data can generally be improved by correcting for imbalanced data.