Modeling and prediction of feeding links using trait data

Background/Question/Methods

Food webs, networks of feeding links in an ecosystem, serve as a basic tool for understanding population dynamics, extinctions, and species distributions. Now, using new, high-resolution data sets and improved computational methods, it is increasingly possible to quantitatively test specific hypotheses about food-web structure and perform rich data analysis on food webs. To date, the focus of the analysis of food-web structure has centered on flexible generative models and the identification of broad patterns. However, in order to better understand specific webs, a discriminative approach, where the presence of feeding links is predicted from species traits, should also be useful. Additionally, by focusing on predictive ability, discriminative models may be able to adaptively guide sampling efforts.

The food web of the Weddell Sea in Antarctica represents a new generation of high-resolution food-web data, and includes species in several habitat zones that range over many orders of magnitude in size. The focus of this study is to examine to what extent links in the Weddell Sea food web can be predicted from species traits, including continuous traits (e.g., body size) and categorical variables (e.g., metabolic class, habitat type), via several different models: (1) a logistic regression model based only on continuous traits; (2) logistic regressions between pairs of groups determined from matching rules for categorical variables; and (3) a support vector machine (SVM), a standard technique from machine learning.

Results/Conclusions

Models based on species trait data provide substantially better predictive ability over simple null generative models. Continuous traits alone do barely better than a random graph model, but the inclusion of categorical variables as predictors accounts for an order of magnitude more links. The standard machine-learning technique, the support vector machine, delivers the best predictive ability, but the matching rule-based model performs reasonably well, and groups species in a straightforward, biologically interpretable way.