Networks are a powerful way to represent the complexity of large ecological systems. However, most ecological networks, such as food webs, contain only partial lists of species interactions. Computational methods for inferring missing links can facilitate field work and investigations of ecological processes. Here, we describe a stacked generalization approach to predict missing links in food webs that accounts for ecological assumptions including link direction. Tests of this method on synthetic food webs show that it can learn to optimally combine structural and trait-based predictions. On a global database of 290 food webs, the method often achieves near-perfect performance, performs better when it can exploit both species traits and network structure, and is principally driven by a subset of ecologically-interpretable predictors. Furthermore, we find that link predictability varies with ecosystem and network characteristics. These results show broad applicability of stacked generalization for predicting and understanding ecological interactions.
