Covariate-informed link prediction with extreme taxonomic bias
Abstract
Biotic interactions provide a valuable window into the inner workings of complex ecological communities and capture the loss of ecological function often precipitated by environmental change. However, the financial and logistical challenges associated with collecting interaction data result in networks that are recorded with geographical and taxonomic bias, particularly when studies are narrowly focused. We develop an approach to reduce bias in link prediction in the common scenario in which data are derived from studies focused on a small number of species. Our Extended Covariate-Informed Link Prediction (COIL+) framework utilizes a latent factor model that flexibly borrows information between species and incorporates dependence on covariates and phylogeny, and introduces a framework for borrowing information from multiple studies to reduce bias due to uncertain species occurrence. Additionally, we propose a new trait matching procedure which permits heterogeneity in trait-interaction propensity associations at the species level. We illustrate the approach through an application to a literature compilation data set of 268 sources reporting frugivory in Afrotropical forests and compare the performance with and without correction for uncertainty in occurrence. Our method results in a substantial improvement in link prediction, revealing 5,255 likely but unobserved frugivory interactions, and increasing model discrimination under conditions of great taxonomic bias and narrow study focus. This framework generalizes to a variety of network contexts and offers a useful tool for link prediction given networks recorded with bias.