Pollen specialist bee species are accurately predicted from visitation, occurrence and phylogenetic data

An animal’s diet breadth is a central aspect of its life history. Yet information about which species have narrow dietary breadths (specialists) and which have comparatively broad dietary breadths (generalists) is missing for many taxa and regions. One possible way to address this gap is to leverage interaction data found on museum specimens and published in the literature. Here, we use bees as our focal taxon to predict dietary specialization and generalization using machine learning models and interaction data, along with a bee phylogeny, and occurrence data for 682 bee species native to the United States. To assess whether our models can transfer to new regions or taxa, we used spatial and phylogenetic blocking in assessing model performance. We found that specialist bees mostly visit their host plants, and that they can be predicted with high accuracy (mean 92% accuracy). Overall model performance was high (mean AUC = 0.84), and our models did a moderate job of predicting generalist bee species, the minority class in our dataset (mean 62% accuracy). Models tested on spatially and phylogenetically blocked data had comparable performance to models tested on randomly blocked data. Our results suggest it is possible to predict specialist bee species in regions and for taxonomic groups where they are unknown but it may be more challenging to predict generalists. Researchers looking to identify pollen specialist and generalist species can generate candidate lists of these species by training models on bees from nearby regions or closely related taxa.