DynaOD: Dynamic Origin-Destination Flow Generation with Discrete-to-Continuous Temporal Semantic Modeling

DynaOD is a machine learning framework that generates realistic urban mobility patterns by modeling temporal dynamics through discrete directional trends and continuous evolution, without requiring historical origin-destination data. The approach uses semantic temporal signals to condition pretrained OD generators, achieving better accuracy and distributional fidelity than existing methods with cross-city transferability.

DynaOD addresses a fundamental challenge in urban computing: synthesizing realistic mobility patterns from temporal context alone. Traditional origin-destination flow prediction relies heavily on historical observations, which limits applicability in new cities or during unprecedented conditions. This research decouples temporal semantics into two complementary components—discrete directional shifts that capture qualitative changes in activity patterns and continuous temporal evolution that models how these shifts unfold—enabling more nuanced representation of urban dynamics.

The framework's modular architecture represents a meaningful advancement in how temporal intelligence can augment existing spatial models. By conditioning pretrained static generators with time-varying region representations, DynaOD achieves scalability without retraining entire systems. This plug-and-play approach has practical implications for smart city infrastructure, ride-sharing platforms, and urban planning applications that must operate across diverse geographic contexts.

From an implementation perspective, the method's cross-city transferability addresses a critical gap in mobility modeling. Urban systems exhibit regional variations that transfer learning struggles to capture; DynaOD's semantic approach appears to generalize more effectively across different city characteristics. The public code release democratizes access to this technique for researchers and practitioners.

For developers building location-based services or urban analytics platforms, this framework offers a pathway to improve flow prediction accuracy without extensive local data collection. The approach could enhance resource allocation in logistics, reduce congestion modeling errors, and improve emergency response planning. The consistent outperformance over baselines suggests practical deployment potential in production environments.

• →DynaOD generates origin-destination flows using only temporal semantics without historical mobility data, enabling deployment in new or data-sparse cities.
• →The dual modeling of discrete directional trends and continuous temporal evolution captures urban activity pattern shifts more accurately than prior methods.
• →The modular, plug-and-play design enables lightweight integration with existing OD generators and supports cross-city transfer learning.
• →Public code availability accelerates adoption among urban computing researchers and location-based service developers.
• →Demonstrated improvements in both predictive accuracy and distributional fidelity position the method for practical deployment in smart city and logistics applications.