Contextual evidence for categorising interactions and guiding their discovery

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

Ecological interactions shape community dynamics and stability, yet many go unrecorded. Link prediction methods attempt to tackle incomplete knowledge, but predictions are only conjectures, and validating them all in the field is neither feasible nor desirable. Furthermore, interactions differ in the action needed for their detection. Validation therefore must be guided by ecological and statistical theory. We present a guided-sampling framework that combines predictions with local and regional observations, using within-system variability across replicated networks as contextual evidence. It assigns each link to a fine-resolution, ecologically-aware taxonomy that separates conflated categories (e.g., forbidden versus undersampled interactions). Each category then implies a concrete action (e.g., sample more, change method). A Bayesian formulation quantifies confidence in each assignment and incorporates researchers' knowledge. We include a worked empirical example, and an interactive Bayesian version online (http://lpguide.ecomplab.com). Contextual evidence turns link prediction from a source of untested hypotheses into a plan for fieldwork.