The major-element chemical composition of garnet provides valuable petrogenetic information, particularly in metamorphic rocks. When facing detrital garnet, information about the bulk-rock composition and mineral paragenesis of the initial garnet-bearing host-rock is absent. This prevents the application of chemical thermo-barometric techniques and calls for quantitative empirical approaches. Here we present a garnet host-rock discrimination scheme that is based on a random forest machine-learning algorithm trained on a large dataset of 13,615 chemical analyses of garnet that covers a wide variety of garnet-bearing lithologies. Considering the out-of-bag error, the scheme correctly predicts the original garnet host-rock in (i) > 95% concerning the setting, that is either mantle, metamorphic, igneous, or metasomatic; (ii) > 84% concerning the metamorphic facies, that is either blueschist/greenschist, amphibolite, granulite, or eclogite/ultrahigh-pressure; and (iii) > 93% concerning the host-rock bulk composition, that is either intermediate–felsic/metasedimentary, mafic, ultramafic, alkaline, or calc–silicate. The wide coverage of potential host rocks, the detailed prediction classes, the high discrimination rates, and the successfully tested real-case applications demonstrate that the introduced scheme overcomes many issues related to previous schemes. This highlights the potential of transferring the applied discrimination strategy to the broad range of detrital minerals beyond garnet. For easy and quick usage, a freely accessible web app is provided that guides the user in five steps from garnet composition to prediction results including data visualization.