A Community Detection Method to Counter the Semantic Noise of Complex Networks: Bridging a Topology and Semantic Subspace Transformation

Abstract

Community detection of complex networks helps understand the network’s composition and explores the behavior characteristics among users. The methods based on nonnegative matrix factorization (NMF) for community detection have been widely studied and applied because of their flexibility and high interpretability. Nevertheless, accompanied by ideal assumptions, classic community detection methods face the following problems. First, they lead to a mismatch between communities, i.e., topology and semantic information. Second, they cannot overcome the effect of macroscopic semantic noise on methods. Finally, they cannot retain the nodes’ distance information in the high-dimensional topology space. Aiming at the these problems, this article proposes a community detection method with self-perception ability that solves the problem of macro semantic noise. The aforementioned issues are overcome by introducing the subspace transfer module and the generalized Laplacian self-perception module. We understand the network factors that affect the model’s performance through empirical observation. Extensive experiments on six representative methods on several benchmark datasets support our conclusion and demonstrate that our method outperforms the representative methods in many datasets.