An appraisal of graph embeddings for comparing trophic network architectures

Christophe Botella (1) , Stéphane Dray (2) , Catherine Matias (3) , Vincent Miele (2) , Wilfried Thuiller (1)

1 LECA - Laboratoire d'Ecologie Alpine
2 LBBE - Laboratoire de Biométrie et Biologie Evolutive - UMR 5558
3 LPSM (UMR_8001) - Laboratoire de Probabilités, Statistique et Modélisation

Comparing the architecture of interaction networks in space or in time is crucial to understand community assembly, trajectory, functioning and persistence. Graph embeddings, that map networks into a vector space where close networks have similar architectures, might represent ideal tools for that purposes. Here, we evaluated the capacity of seven graph embedding approaches to disentangle the architectural similarities of interactions networks for posterior supervised and unsupervised analytic tasks. The evaluation was carried out over a large number of simulated trophic networks representing variations around six trophic architectural properties and sizes. We did not find an overall best approach and instead showed that the performance of the approaches depended on the targeted architectural properties and thus on the research questions. We also highlighted the importance of a network size normalization of the embedding to avoid meaningless variability blurring posterior unsupervised analysis. We concluded by orientating potential users to the most suited approaches given the question, the targeted architectural property, and outlined links between those properties and three ecological processes: robustness to extinction, community persistence and ecosystem functions. This study should enhance the appropriation of graph embedding approaches by ecologists.

Mots clés: dimension reduction; ecological interaction networks; evaluation; food-webs; graph embedding; species interactions; trophic networks; trophic groups

Domaines: Biodiversité et Ecologie; Machine Learning [stat.ML]; Applications [stat.AP]; Méthodologie [stat.ME]; Sciences de l'environnement; Mathématiques [math]; Statistiques [stat]

Christophe Botella, Stéphane Dray, Catherine Matias, Vincent Miele, Wilfried Thuiller. An appraisal of graph embeddings for comparing trophic network architectures. 2021. ⟨hal-03191630v1⟩