Predicting cryptic links in host-parasite networks

Networks are a way to represent interactions among one (e.g., social networks) or more (e.g., plant-pollinator networks) classes of nodes. The ability to predict likely, but unobserved, interactions has generated a great deal of interest, and is sometimes referred to as the link prediction problem. However, most studies of link prediction have focused on social networks, and have assumed a completely censused network. In biological networks, it is unlikely that all interactions are censused, and ignoring incomplete detection of interactions may lead to biased or incorrect conclusions. Previous attempts to predict network interactions have relied on known properties of network structure, making the approach sensitive to observation errors. This is an obvious shortcoming, as networks are dynamic, and sometimes not well sampled, leading to incomplete detection of links. Here, we develop an algorithm to predict missing links based on conditional probability estimation and associated, node-level features. We validate this algorithm on simulated data, and then apply it to a desert small mammal host-parasite network. Our approach achieves high accuracy on simulated and observed data, providing a simple method to accurately predict missing links in networks without relying on prior knowledge about network structure.Author summary: The majority of host-parasite associations are poorly understood or not known at all because the number of associations is so vast. Further, interactions may shift seasonally, or as a function of changing host densities. Consequently, host-parasite networks may be poorly characterized since effects of cryptic host-parasite associations on network structure are unknown. To address this, we developed theory and applied it to empirical data to test the ability of a simple algorithm to predict interactions between hosts and parasites. The algorithm uses host and parasite trait data to train predictive probabilistic models of host-parasite interaction. We tested the accuracy of our approach using simulated networks that vary greatly in their properties, demonstrating high accuracy and robustness. We then applied this algorithm to data on a small mammal host-parasite network, estimated model accuracy, identified host and parasite traits important to prediction, and quantified expected changes to structural properties of the network as a result of link relabeling.