Tailoring Benchmark Graphs to Real-World Networks for Improved Prediction of Community Detection Performance

Analysts interested in understanding the community structure of a particular real-world network will often simply choose a popular community detection algorithm and trust the generated results without further investigation, but algorithm performance can vary depending on the network characteristics. We demonstrate that by running experiments on benchmark graphs tailored to match characteristics of a real-world network of interest, a better understanding can be obtained on how community detection algorithms will perform on the real-world network. We show that the correlation between the performance of the community detection methods on a publicly available dataset to the average performance of the same methods on the corresponding tailored benchmark graphs is high whereas the correlation with LFR benchmark graphs is negative. This means the methods that performed well on the tailored graphs also performed well on the real-world network but methods that perform well on LFR graphs did not perform well on the real-world network, demonstrating that the proposed methodology has merit.