Comparison of methods to identify modules in noisy or incomplete brain networks

Community structure, or "modularity," is a fundamentally important aspect in the organization of structural and functional brain networks, but their identification with community detection methods is confounded by noisy or missing connections. Although several methods have been used to account for missing data, the performance of these methods has not been compared quantitatively so far. In this study, we compared four different approaches to account for missing connections when identifying modules in binary and weighted networks using both Louvain and Infomap community detection algorithms. The four methods are "zeros," "row-column mean," "common neighbors," and "consensus clustering." Using Lancichinetti-Fortunato-Radicchi benchmark-simulated binary and weighted networks, we find that "zeros," "row-column mean," and "common neighbors" approaches perform well with both Louvain and Infomap, whereas "consensus clustering" performs well with Louvain but not Infomap. A similar pattern of results was observed with empirical networks from stereotactical electroencephalography data, except that "consensus clustering" outperforms other approaches on weighted networks with Louvain. Based on these results, we recommend any of the four methods when using Louvain on binary networks, whereas "consensus clustering" is superior with Louvain clustering of weighted networks. When using Infomap, "zeros" or "common neighbors" should be used for both binary and weighted networks. These findings provide a basis to accounting for noisy or missing connections when identifying modules in brain networks.