COS 37-1 - Predicting missing interactions in food webs using stacked models and species traits

Abstract: Ecosystems currently face multiple human impacts including habitat loss, climate change, and the introduction of invasive species, leading to biodiversity loss and threats to ecosystem services. Species rely on one another to survive, so if a disturbance affects the abundance of one species, other species can also be affected. One way to predict and understand the impacts of anthropogenic disturbances on biodiversity and ecosystem services is to use species interaction networks, such as food webs, to depict and model these patterns of species interdependency. However, collecting data on species interactions is labor-intensive so accounts of species interactions remain incomplete in many ecosystems. In this work, we develop and apply an approach for predicting missing interactions in ecological networks that draws on recent advances in network science. Rather than using a single model to predict missing interactions in ecological networks, our approach leverages model stacking to combine a diverse ensemble of prediction methods. Each individual interaction prediction method tries to distinguish between missing interactions and true negatives, and the stacking approach learns how to combine these individual methods to make the best interaction predictions in different contexts. We adapt this approach to the context of directed, attributed networks and evaluate missing interaction prediction performance across a global food web database. Attributed networks are those in which each node has associated metadata, so this adaptation allows us to consider available species-level information, in this case species body mass, metabolic type, and movement type.

We compare the performance across the database of three stacked models: one which uses only species attributes (e.g., traits), one which uses only network structure (e.g., predictions based on species degree in the network), and one which combines species attributes and structure for prediction. We find that all three stacked models perform better than random at identifying missing interactions across the food web database. Further, the model combining aspects of network structure and species attributes performs the best on average. These results demonstrate that both species traits and observed species interaction network structure can be used to predict missing interactions in food webs, but that one will often get the best results if species traits and network structure are considered together. This work demonstrates a flexible approach for species interaction prediction in applied contexts, for example for correcting interaction datasets or predicting novel interactions that will be formed by a species introduced into an ecosystem.