The inherent privacy risks associated with trajectory data limit its applicability in domains such as transportation planning and urban management. Generating data that accurately reflects real-world human mobility patterns while ensuring rigorous privacy protection has become an imperative challenge. Existing trajectory generation approaches primarily model individual movement behaviors, neglecting the integration of individual and collective mobility characteristics, which often leads to synthetic trajectories lacking authenticity. To overcome these limitations, we propose SyncMove, a novel trajectory generation framework designed to synchronize individual mobility patterns with underlying collective dynamics, thereby enhancing authenticity without compromising privacy. Specifically, we leverage a temporal graph contrastive learning mechanism to capture collective-level mobility patterns from check-in sequences, effectively modeling both temporal and structural dependencies in collective behavior. These collective features are then fused with individual mobility representations and fed into a generative adversarial network for trajectory generation. Crucially, to balance the learning of these two aspects, we design a progressive user loss that prioritizes collective patterns early in training before gradually shifting focus to individual heterogeneity. This staged approach ensures that generated trajectories are grounded in collective mobility patterns while retaining individual uniqueness. By learning the underlying data distribution without directly storing real user records, SyncMove inherently preserves privacy. Extensive experiments on real-world datasets from New York and Singapore demonstrate that SyncMove consistently outperforms baseline methods across multiple metrics, yielding realistic synthetic trajectories without compromising user privacy.
Li et al. (Thu,) studied this question.