Temporal link and unlink prediction using multiscale network dynamics

Link and Unlink prediction play a vital role in comprehending network dynamics. To precisely predict links and unlinks, multiscale information sources in dynamic networks are vital, involving the simultaneous exploitation of microscopic higher-order information, mesoscopic community information, and temporal information of the dynamic network. However, most existing approaches analyze these factors independently, leading to limited predictive performance. To address this limitation, this study proposes a novel model, namely LUComHT, for forecasting link formation and removal with microscopic, mesoscopic, and temporal information on dynamic networks. LUComHT comprises three key steps: for each network snapshot, LUComHT first exploits microscopic higher-order network structure using a local random walk method and applies a Louvain algorithm to extract mesoscopic community information. Then, LUComHT initializes and optimizes a sequence of temporal matrices and a global matrix for reconstructing the microscopic information and mesoscopic information of the snapshots using a non-negative matrix factorization (NMF) approach based on the microscopic and mesoscopic information of the snapshots. Lastly, LUComHT computes the future similarity matrix of the snapshots to forecast link formation and disappearance. Extensive experiments on multiple real-world dynamic networks demonstrate that LUComHT consistently outperforms existing methods, highlighting the importance of multiscale information fusion. Furthermore, by jointly modeling these multiscale information sources, LUComHT represents a significant methodological advancement, offering a more holistic and accurate characterization of network evolution. The findings also provide policymakers with a data-driven framework to anticipate changes in social, institutional, and digital connectivity, enabling evidence-based strategies that strengthen collaboration networks, improve governance, and support resilient systems.