transfer learning ancestral character estimation biogeography 1. Despite their importance in many ecological processes, collecting data and information on ecological interactions is an exceedingly challenging task. For this reason, large parts of the world have a data deficit when it comes to species interactions, and how the resulting networks are structured. As data collection alone is unlikely to be sufficient, community ecologists must adopt predictive methods.2. We present a methodological framework that uses graph embedding and transfer learning to assemble a predicted list of trophic interactions of a species pool for which their interactions are unknown. Specifically, we 'learn' the information (latent traits) of species from a known interaction network and infer the latent traits of another species pool for which we have no a priori interaction data based on their phylogenetic relatedness to species from the known network. The latent traits can then be used to predict interactions and construct an interaction network.3. Here we assembled a metaweb for Canadian mammals derived from interactions in the European food web, despite only 4% of common species being shared between the two locations. The results of the predictive model are compared against databases of recorded pairwise interactions, showing that we correctly recover 91% of known interactions.4. The framework itself is robust even when the known network is incomplete or contains spurious interactions making it an ideal candidate as a tool for filling gaps when it comes to species interactions. We provide guidance on how this framework can be adapted by substituting some approaches or predictors in order to make it more generally applicable.

Despite their importance in many ecological processes, collecting data and information on ecological interactions, and therefore species interaction networks, is an exceedingly challenging task. For this reason, large parts of the world have a deficit of data of which species interact, and what we can expect the network structure of these interactions to be. As data collection alone is unlikely to be sufficient at filling these global gaps, community ecologists must adopt predictive methods. In this contribution we develop such a method, relying on graph embedding (the extraction of explanatory latent variables from known graph structures) and transfer learning (the application of previous solution to novel problems with limited predictors overlap) in order to assemble a predicted list of trophic interactions between mammals of Canada. This interaction list is derived from extensive knowledge of the mammalian food web of Europe, despite the fact that there are fewer than 5% of common species between the two locations. We provide guidance on how this method can be adapted by substituting some approaches or predictors in order to make it more generally applicable to a broad family of ecological problems.

Metawebs, i.e. networks of potential interactions within a species pool, are a powerful abstraction to understand how large-scales species interaction networks are structured.Because metawebs are typically expressed at large spatial and taxonomic scales, assembling them is a tedious and costly process; predictive methods can help circumvent the limitations in data deficiencies, by providing ‘draft’ metawebs.One way to improve the predictive ability is to maximize the information used for prediction, by using graph embeddings rather than the list of species interactions. Graph embedding is an emerging field in machine learning that holds great potential for ecological problems.In this perspective, we outline how the challenges associated with inferring metawebs line-up with the advantages of graph embeddings; furthermore, because metawebs are inherently spatial objects, we discuss how the choice of the species pool has consequences on the reconstructed network, but also embeds hypotheses about which human-made boundaries are ecologically meaningful.