30 October 2024 in Knowledge-Based Systems
10 April 2026 in SSRN Electronic Journal
Cellular signaling data continuously and extensively records user mobility behaviors, which is important for urban computing. Meanwhile, such data is characterized by large-scale event streams, continuous temporal dynamics and strong spatial semantics, imposing higher requirements on representation learning methods. Existing methods still have several limitations in this scenario, including limited scalability on city-scale data, separate modeling of user and area representations, and insufficient fusion of temporal and spatial information.To address these issues, we propose MoDyGEM, an event-driven framework for joint user-area embedding from large-scale cellular signaling streams. MoDyGEM models signaling records as a continuous-time bipartite dynamic graph of users and areas. For each event, it constructs a user-centered local subgraph and an area-centered local subgraph to capture the context. Then, it jointly encodes node, temporal, and spatial information into a unified interaction representation. To further aggregate local contextual signals a subgraph mixer is developed. A residence-time-weighted feature updater is introduced to incrementally maintaining global dynamic embeddings.For training, MoDyGEM is pretrained with self-supervised temporal link prediction, and the learned representations are transferred to both user-side and area-side downstream tasks. Experiments on real-world Beijing cellular signaling data show that MoDyGEM outperforms multiple strong baselines on various tasks.