GTMNet: a generalized task-oriented memory network for urban mobility forecasting

Accurate urban mobility forecasting plays a vital role in supporting informed decision-making for traffic management and policy design. However, existing approaches continue to face key challenges in balancing accuracy with computational efficiency, scaling to large mobility datasets, and ensuring transferability across diverse contexts. To address these limitations, we propose the generalized task-oriented memory network (GTMNet), a unified framework for lightweight and robust multi-task forecasting. GTMNet comprises three core components: (i) a memory-augmented encoder for continuous-time multi-source data representation, (ii) a spatiotemporal transformer (STFormer) module for capturing fine-grained spatiotemporal dependencies, and (iii) a task-specific mixture-of-experts architecture for task-aware prediction. To further enhance efficiency and robustness, GTMNet integrates a two-stage query mechanism that selectively extracts task-specific historical interactions, enabling fine-grained modeling with minimal computational overhead. Additionally, message-level dropout and adaptive low-rank decomposition are introduced to mitigate overfitting and promote robust generalization, particularly under high-dimensional and sparse origin–destination demand conditions. Extensive experiments on four large-scale mobility datasets across countries and systems demonstrate that GTMNet consistently outperforms state-of-the-art baselines in accuracy and efficiency. Ablation studies and case analyses further validate the role of each module and reveal interpretable mobility patterns across varying spatiotemporal contexts, underscoring GTMNet’s scalability and practical potential for real-world urban mobility forecasting. These findings highlight both the theoretical contributions and the practical potential of GTMNet for real-world applications such as dynamic traffic prediction, demand-responsive transit planning, and policymaking.