Contextual evidence for categorising interactions and guiding their discovery

Kesem Abramov, Shir Miriam Nehoray, Rami Puzis, Peter Mucha, Shai Pilosof 

A surge of recent methods tackles incomplete ecological interaction data through link prediction. However, validating predictions means field-sampling an immense number of candidate interactions, possibly with a poorly suited method. This costly endeavour lacks operative guides grounded in strong ecological and statistical theory. We present a guided-sampling framework combining link prediction with within-system variability across replicated networks as contextual evidence, generating a fine-resolution, ecologically-aware link taxonomy. It resolves categories ecologists have long struggled to separate, such as forbidden interactions versus those merely missing from the local sample, pinpoints the most probable unknowns, and directs targeted, cost-efficient discovery. We formalise how confidence in each assignment is quantified and grows as evidence accumulates. Treating validation as evidence accumulation makes it a tractable sampling problem rather than a permanent limit on knowledge. We provide a worked analysis of an empirical data set and an online guide at http://lpguide.ecomplab.com.