Identification of dynamic community in temporal network via joint learning graph representation and nonnegative matrix factorization

摘要

The accumulated temporal networks provide an opportunity to explore the dynamics of complex systems, and detecting the dynamic communities is of considerable significance because they shed light on revealing the mechanisms of the underlying systems. In comparison to static communities, identifying dynamic community must simultaneously take into account the clustering accuracy and drift, which are challenging to balance. Furthermore, the available algorithms cannot fully characterize the dynamics of networks, thereby resulting in undesirable performance. To attack these issues, we propose a novel joint Learning of Dynamic Embedding and Clustering (jLDEC) algorithm for dynamic community detection, where network embedding, dynamics of edges, and clustering are integrated into an overall objective function. Specifically, the graph representation and dynamics at the edge level are fused, which provide a better way to characterize the dynamics of communities. The joint learning framework extracts the graph representation under the guidance of clustering, which promotes the accuracy of clustering in return. The experimental results on the artificial and real-world networks demonstrate that the proposed method outperforms state-of-the-art approaches for dynamic community detection in temporal networks.The accumulated temporal networks provide an opportunity to explore the dynamics of complex systems, and detecting the dynamic communities is of considerable significance because they shed light on revealing the mechanisms of the underlying systems. In comparison to static communities, identifying dynamic community must simultaneously take into account the clustering accuracy and drift, which are challenging to balance. Furthermore, the available algorithms cannot fully characterize the dynamics of networks, thereby resulting in undesirable performance. To attack these issues, we propose a novel joint Learning of Dynamic Embedding and Clustering (jLDEC) algorithm for dynamic community detection, where network embedding, dynamics of edges, and clustering are integrated into an overall objective function. Specifically, the graph representation and dynamics at the edge level are fused, which provide a better way to characterize the dynamics of communities. The joint learning framework extracts the graph representation under the guidance of clustering, which promotes the accuracy of clustering in return. The experimental results on the artificial and real-world networks demonstrate that the proposed method outperforms state-of-the-art approaches for dynamic community detection in temporal networks.(c) 2021 Elsevier B.V. All rights reserved.