COS 259-1 - CANCELLED - Making something out of nothing at all: Food web reconstruction using transfer learning

Thursday, August 18, 2022

1:30 PM – 1:45 PM EDT

Location: 518A

Presenting Author(s)

TS

Tanya Strydom

University of Montreal, Canada

Coauthor(s)

SB

Salomé Bouskila

University of Montreal, Canada
Francis Banville, n/a

PhD candidate
Université de Montréal
Montréal, Quebec, Canada
CB

Ceres Barros

University of British Columbia, Canada
DC

Dominique Caron

PhD Student
McGill University, Quebec, Canada
MF

Maxwell Farrell

University of Toronto, Canada
MF

Marie-Josée Fortin

University of Toronto, Ontario, Canada
VH

Victoria Hemming, PhD

Postdoctoral Fellow
The Conservation Decisions Lab, Department of Forest and Conservation Sciences, University of British Columbia, Canada
BM

Benjamin Mercier

Université de Sherbrooke, Canada
LP

Laura Pollock

Assistant Professor
McGill University, Quebec, Canada
RR

Rogini Runghen

University of Canterbury, New Zealand
GD

Giulio Dalla Riva

University of Canterbury, New Zealand
TP

Timothée Poisot, PhD

Associate Professor
Université de Montréal, Quebec, Canada

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.