Unveiling the hidden species interactions in a mutualistic seed-dispersal network through the use of machine learning classifiers.

Due to anthropogenic pressures such as climate change, natural communities will alter their geographical range to accomadate the shifting climate. This will lead to network rewiring and a change species interactions. Currently, we do not have complete information of all realized interactions in species communities, let alone the future novel species communities. The lack of knowledge is a major hinderance to assess the vulnerability of both the current and future species communities to perturbations.<br/><br/>Here we present a method based on machine learning, to infer potential – but unobserved – species interactions. We trained a machine learning classifier to predict mutualistic interactions between Afrotropical figs (Ficus) and their seed-dispersing frugivores on a database of 4807 observed interactions, the functional traits of the plants and animals and phylogenetic relationships. We simulated non-random extinction scenarios using observed versus predicted interactions to evaluate extinction robustness across regional network ensembles.<br/><br/>Our machine-leaning model trained on observed interactions, traits and phylogeny was able to classify interactions with high fidelity. We found distinct spatial patterns in the number and nature of inferred interactions, and that using only observed interactions underestimates the robustness towards non-random extinction of local networks when compared to using inferred potential interactions. Predicting trophic interactions can thus provide fundamental insights how climate change and loss of keystone species can have unforeseen cascading effects on community stability.<br/>