Summary Shifts among functional pollinator groups are commonly regarded as sources of floral morphological diversity (disparity) through the formation of distinct pollination syndromes. While pollination syndromes may be used for predicting pollinators, their predictive accuracy remains debated, and they are rarely used to test whether floral disparity is indeed associated with pollinator shifts. We apply classification models trained and validated on 44 functional floral traits across 252 species with empirical pollinator observations and then use the validated models to predict pollinators for 159 species lacking observations. In addition, we employ multivariate statistics and phylogenetic comparative analyses to test whether pollinator shifts are the main source of floral disparity in Melastomataceae. We find strong support for four well‐differentiated pollination syndromes (‘buzz‐bee’, ‘nectar‐foraging vertebrate’, ‘food‐body‐foraging vertebrate’, ‘generalist’). While pollinator shifts add significantly to floral disparity, we find that the most species‐rich ‘buzz‐bee’ pollination syndrome is most disparate, indicating that high floral disparity may evolve without pollinator shifts. Also, relatively species‐poor clades and geographic areas contributed substantially to total disparity. Finally, our results show that machine‐learning approaches are a powerful tool for evaluating the predictive accuracy of the pollination syndrome concept as well as for predicting pollinators where observations are missing.