Application and comparative performance of network modularity algorithms to ecological communities classification

Network modularity is a well-studied large-scale connectivity pattern in networks. The detection of modules in real networks constitutes a crucial step towards a description of the network building blocks and their evolutionary dynamics. The performance of modularity detection algorithms is commonly quantified using simulated networks data. However, a comparison of the modularity algorithms utility for real biological data is scarce. Here we investigate the utility of network modularity algorithms for the classification of ecological plant communities. Plant community classification by the traditional approaches requires prior knowledge about the characteristic and differential species, which are derived from a manual inspection of vegetation tables. Using the raw species abundance data we constructed six different networks that vary in their edge definitions. Four network modularity algorithms were examined for their ability to detect the traditionally recognized plant communities. The use of more restrictive edge definitions significantly increased the accuracy of community detection, that is, the correspondence between network-based and traditional community classification. Random-walk based modularity methods yielded slightly better results than approaches based on the modularity function. For the whole network, the average agreement between the manual classification and the network-based modules is 76% with varying congruence levels for different communities ranging between 11% and 100%. The network-based approach recovered the known ecological gradient from riverside - sand and gravel bank vegetation - to dryer habitats like semidry grassland on dykes. Our results show that networks modularity algorithms offer new avenues of pursuit for the computational analysis of species communities.