CANCELLED - Making something out of nothing at all: Food web reconstruction using transfer learning

Background/Question/Methods

Collecting data and information on ecological interactions is an exceedingly challenging task. This results in patchily distributed information with large parts of the world having a data deficit when it comes to species interactions, which poses a considerable challenge to understanding the emergent properties of ecological communities. As data collection alone is unlikely to be sufficient, community ecologists must adopt predictive methods to construct metawebs (potential interactions for a region) from which to work with and help in circumventing data limitations. Here we present a methodological framework that uses graph embedding and transfer learning to assemble a predicted metaweb for terrestrial Canadian mammals using the known European food web. Specifically, we ‘learn’ the information (latent traits) of species from Europe and infer the latent traits for the Canadian species pool (for which we have no a priori interaction data) based on their phylogenetic relatedness to species from Europe. The latent traits are inferred using a singular value decomposition (SVD) of the European network, which are then ‘transferred’ to the Canadian species by mapping them onto a reference phylogeny. These latent traits are then used to construct a Canada-level metaweb using a random dot product graph (RDPG) to predict probabilistic interactions.

Results/Conclusions

Despite having a low degree of overlap (only 4%) between European and Canadian species pools, the resulting shape of the predicted Canadian metaweb is ecologically plausible, making biological sense with regards to both trophic organisation and predator-prey body mass relationships. Furthermore, the resulting interactions were drawn from the predictive model and compared against databases of recorded pairwise interactions within Canada, which showed that we were able to correctly recover 91% of known, observed interactions for the region. It should be noted that the specific steps of the transfer learning framework that we present and use are themselves amenable to change, e.g. using a different (situationally more suitable) medium of transfer such as species traits or alternative methods of graph embedding, which lends itself applicability in a broad range of potential scenarios. Additionally, the probabilistic nature of the predicted metaweb may serve as an informative prior when using network inference techniques that are based on Bayesian approaches. Overall transfer learning presents a viable option for predicting species interactions where local data are missing or incomplete through construction of ‘draft’ metawebs and is both computationally inexpensive as well as flexible and easy to implement.