Clustering methods differ in their ability to detect patterns in ecological networks

Abstract

Summary Network ecology has been an extraordinarily fertile field of research over the last 20 years. Its ultimate goal is to understand how the complex systems of interdependent species assemble, function and evolve. Here, we aimed to help ecologists to select the best methods for detecting subgroups of highly interacting species (usually referred to as compartments or modules) in bipartite networks (e.g. plant–pollinator networks, host–parasite networks), because these subgroups may reveal the processes underlying the assembly of the network and may influence its stability. We simulated several thousand bipartite ecological networks and we compared seven methods of network clustering in terms of their ability to retrieve the number and the composition of species subgroups. Among the seven methods compared, we found that the edge‐betweenness algorithm was the best option for binary networks. The stochastic block model was the best method for weighted networks. Modularity maximization, the most popular clustering method in ecology, was among the three best methods in both cases. We thus provide ecology researchers with precise advice concerning the best choice of network clustering method, according to the type of data collected. We also provide the code for simulating bipartite networks and clustering them, in order to facilitate future methodological comparisons.