Graph embedding and transfer learning can help predict species interaction networks despite data limitations

Abstract

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.

Spatially explicit predictions of food web structure from regional level data

“…Our approach presents a way to downscale a metaweb, produce localized predictions using probabilistic networks as inputs and outputs, and incorporate uncertainty, as called for by Strydom et al (2022b). It gives us an idea of what local metawebs or networks could look like in space, given species distributions and their variability, as well as the uncertainty around species interactions.…”

“…In systems where in situ interaction and complete network data are available, the approach put forward by Gravel et al (2019) achieves a similar purpose as we attempted here, but is more rigourous and allows modelling the effect of the environment on the interactions themselves. Without such data, establishing or predicting the metaweb should be the first step toward producing localized predictions (Strydom et al 2022b). Welldocumented binary metawebs such as the European tetrapod metaweb could be partly combined with our approach if used with probabilistic SDMs and summarized by ecoregions (as they would only lack an initial probabilistic metaweb, but would still obtain a more probabilistic output).…”

“…1). The metaweb sets an upper bound on the potential interactions (Strydom et al 2022b), therefore, the downscaled metaweb is a refined upper boundary at the local scale that takes into account co-occurrences. It is still potential in nature and differs from a local realization, from which it should have a different structure.…”

“…the local "realization" of the metaweb when combined with species distributions, Poisot et al 2012) reflect local environmental conditions but still retain the signal of the metaweb to which they belong (Saravia et al 2022). Given this, Strydom et al (2022b) defended that predicting the metaweb's structure should be the core goal of predictive network ecology, as if there is a strong link between the metaweb and its local realizations, more accurante predictions of the metaweb will have the potential to bring us closer to producing accurate local (downscaled) predictions. Therefore, establishing or predicting the metaweb should be the first target in systems lacking information about local realizations.…”

“…A potential limitation to previous downscaling approaches is that they assume interactions are constant across space, which ignores well-documented interaction variability, and masks the effect of environmental conditions on interaction realization (Braga et al 2019). In contrast, recent studies argued that seeing interactions as probabilistic (rather than binary) events allows us to account for their variability in space (Poisot et al 2016) and that this should also be reflected at the metaweb level (Strydom et al 2022b). Gravel et al (2019) introduced a probabilistic framework describing how the metaweb can generate local realizations and showed how it could be used for modelling interaction distributions.…”