The Serengeti food web is rapidly emerging as the most highly-resolved terrestrial web. Most people's familiarity with the mammalian species allow us to use it as a strong intuitive test of the structure and "skill" of food-web models. We test a model of network structure based on groups, which need not be specified a priori and can correspond to tightly connected submodules, trophic guilds, or whatever arrangement is best supported by the link structure, allowing us to ask what features of the Serengeti food web are topologically dominant: trophic structure, spatial submodules, or other features. We present a Bayesian Markov-chain Monte Carlo approach for inference and model selection, sampling over the space of species groupings and model parameters.

Results/Conclusions  

The groups identified in the Serengeti web correspond to trophic levels subdivided into habitat type, with one main predator group, two groups each of herbivores and matching plant species, and other groups relating to more subtle properties of the web. In general, the group model can test existing qualitative knowledge about food webs and also suggest new lines of inquiry and more biologically-organized models. One such model is proposed that combines groups with hierarchical orderings of species. Additionally, the Bayesian modeling approach provides richer inferential information and a natural framework for comparing different models.