Dynamic Community Detection via Kalman Filter-Incorporated Non-negative Matrix Factorization

Abstract

Community detection on dynamic undirected network (DUN) is a vital issue in the area of network representation. Note that most existing studies built a detection model on a static network, which is incompatible with a DUN that is dynamically evolving and contains temporal patterns. Aiming at addressing this issue, this paper proposes a kalman filter-incorporated non-negative matrix factorization -based dynamic community detection (KDCD) model. Its main idea is to precisely track the temporal variations of a DUN with the state-transition function of a kalman filter, as well as accurately fit the numerical characteristics of the target network with an alternating least square solver. Empirical studies on three real-world DUNs demonstrate that the proposed KDCD model outperforms state-of-the-art models in achieving highly-accurate dynamic community detection results.