GLFR: A generalized LFR benchmark for testing community detection algorithms

Comparisons between community detection methods are mostly based on their accuracies in recovering the built-in community structure in artificial benchmark networks. Current community detection benchmarks assign a fixed fraction of inter-community links, referred to as the mixing fraction, for every community in the same network. We first show in this paper that the variation in community mixing fractions has different impacts on the performances of different community detection methods that could change the decision to select a particular detecting algorithm. To comprehensively compare community detection methods, we therefore need a benchmark that generates heterogeneous community mixing fractions, which is not currently available. We address this gap by generalizing the state-of-the-art Lancichinetti-Fortunato- Radicchi benchmark to generate networks with heterogeneous community mixing fractions. Using our new benchmark, we can quantify the impact of the variation in community mixing fractions on existing community detection methods and re- evaluate the performance of the detecting algorithms as a function of the heterogeneity among the mixing fractions. Furthermore, we show that the heterogeneous community mixing tests using our generalized benchmark reflect better the performance that would be expected on real networks than the homogeneous community mixing tests using the original benchmark.