Tue, Aug 16, 2022: 9:00 AM-9:15 AM
518B
Background/Question/MethodsThe observed structure of a consumer-resource network consists of only a subset of potential links within that network. The true structure deviates from the metaweb of potential interactions due to stochastic assembly processes, while sampling effort limitations and error mean our observed network likely imperfectly represents this realized structure. Link prediction methods allow us to construct potential metawebs for a given empirical network. These predictions can guide targeted sampling of predicted links between rare species as well as give insight into potential interactions that may occur as species' ranges shift. These interactions may be predicted using a variety of information types; understanding how different kinds of information compare in their ability to predict links between different types of nodes is important. To this end, we compare random-forest classifier models informed by combinations of phylogenetic structure, species traits, and latent network structure in their ability to predict interactions in a diverse network of fruiting plants and frugivorous birds in Brazil's Atlantic forest. After validating model performance, we applied our models to the full interaction network in order to generate predictions of unobserved or unrealized potential interactions.
Results/ConclusionsLatent network structure was the most important determinant of model predictive performance; models informed only by latent structure consistently outperformed those only utilizing phylogenetic or trait information. Incorporating trait or phylogenetic information into latent models had little to no effect on discriminatory power compared to latent-only models (as measured by AUC) but did increase overall accuracy (as measured by root mean square error). Variable importance for models incorporating all three types of information revealed latent features to have the greatest influence, followed by frugivore phylogeny, gape size, and body mass. Conversely, plant traits such as growth form or fruit color comprised the least informative variables for prediction. Together, our findings suggest that existing network structure is important for predicting links in plant-frugivore networks. This pattern may reflect a relative dominance of neutral factors such as species abundance in determining broadscale network structure as opposed to niche differentiation.
Results/ConclusionsLatent network structure was the most important determinant of model predictive performance; models informed only by latent structure consistently outperformed those only utilizing phylogenetic or trait information. Incorporating trait or phylogenetic information into latent models had little to no effect on discriminatory power compared to latent-only models (as measured by AUC) but did increase overall accuracy (as measured by root mean square error). Variable importance for models incorporating all three types of information revealed latent features to have the greatest influence, followed by frugivore phylogeny, gape size, and body mass. Conversely, plant traits such as growth form or fruit color comprised the least informative variables for prediction. Together, our findings suggest that existing network structure is important for predicting links in plant-frugivore networks. This pattern may reflect a relative dominance of neutral factors such as species abundance in determining broadscale network structure as opposed to niche differentiation.