Correlation enhanced modularity-based belief propagation method for community detection in networks

Community structure is an important feature of networks, and the correct detection of communities is a fundamental problem in network analysis. Statistical inference has recently been proposed for successful detection, provided the number of communities can be appropriately estimated a priori. In the absence of such information, model selection by determination of the number of communities remains an issue. We show here that correlation between communities from a highly parceled partition can be used to estimate a narrow range of variation for the real number of communities. This range, further elaborated by modularity-based belief propagation, correctly identifies communities. Testing on synthetic networks generated by a stochastic block model and a set of real-world networks shows that our method can alleviate the effects of modularity fluctuations well and enhance the ability of community detection of the bare modularity-based belief propagation method.