A Comparative Evaluation of Community Detection Algorithms in Social Networks

Community detection in networks is used to understand the underlying structure of the network and obtain insight into the structure of the network. Evaluation of the detected structures is challenging since there are no canonical solutions available. Validation of developed algorithms therefore requires evaluation on multiple different data sets. In this paper, we use synthetic data sets generated by Lancichinetti, Fortunato and Radicchi (LFR) algorithm, and real datasets from known networks to evaluate several different community detection algorithms. The LFR benchmark is used to generate networks with a variety of sizes, and mixing parameters ranging from completely separable to highly mixed for this comparative study. We compare the performance of the recently developed Community Density Rank (CDR) algorithm against other state-of-the-art community detection approaches. It is found that for the synthetic datasets the CDR outperforms the comparison algorithms while the results are mixed for all algorithms on real datasets. These results are used to generalize the applicability of community detection algorithms across a variety of datasets.