Identification of the effects of the existing network properties on the performance of current community detection methods

Abstract

Community detection has attracted many attentions recently. Considering the effect of current network structure on the result of the recent community detection methods is useful to yield a probable performance trade-off for future algorithm selection. In this paper, we first offer a new ranking method with 3 levels for small-world and scale-free networks to measure such properties more accurately, in determining their influences on the methods performance. Thereafter, we examine 12 popular community detection methods and 43 related datasets. The results show that 24 datasets have small-world properties, 5 datasets have scale-free properties, and 9 datasets have both. However, 5 of them have no features of small-world or scale-free networks. It is also observable that 4 methods work better for networks with small-world features and 8 for both small-world and scale free. Finally, we propose a flexible community detection method based on the detected network type.