Food web ecology in space is limited by data availability at broad spatial, environmental, and taxonomic scales. A lot of places are therefore "information deserts", where the volume of data is too small to allow for meaningful inference about the properties, structure, and dynamics of ecological networks. In a context where the vertical diversity of communities is increasingly used as an indicator with potential policy relevance, we have an urgent need to be able to generate information in places where data are lacking. One possible avenue is to rely on transfer learning, namely the encoding of knowledge about one problem in a model of another problem; in space, this can allow the transfer of information from data-rich to data-deficient regions. Yet, the structure of food webs (discrete, sparse) does not always apply to most common machine learning approaches. To circumvent this problem, I present an illustration of a method coupling graph embedding as a dense, continuous network representation and transfer learning, to predict the trophic interactions between mammals in Canada based on a manually curated dataset from Eastern and Central Europe. Despite having only 4% of taxa in common, our method predicts Canadian interactions with about 92% accuracy.