Sensitivity analysis of computational graph-based dynamic multi-modal O–D estimation under data and behavioral uncertainties

This study investigates the sensitivity of multi-modal dynamic origin–destination estimation (DODE) under realistic data and modeling uncertainties. Although DODE has been widely studied for integrating system-level data and dynamic traffic simulation, limited attention has been paid to how estimation outcomes respond to imperfect data coverage, behavioral assumptions, and initialization conditions, especially for generic multi-modal transportation systems. Understanding these sensitivities is critical for practical insights. Building upon a previously developed holistic multi-modal DODE framework, this research performs systematic and scalable sensitivity analyses enabled by a computational graph-based learning approach. The framework jointly estimates dynamic O–D demand for private cars and public transit (bus and metro) by minimizing discrepancies between simulated and observed data, including, but are not limited to, link flow, travel time, and boarding/alighting counts. The computational graph-based learning provides a standardized and flexible framework that can seamlessly integrate and leverage any combination of multi source data, even in the presence of missing or incomplete coverage, enabling a broader range of sensitivity analyses. The simulation-based DODE also enables more refined sensitivity analysis on behavioral or operational factors embedded in the simulation. Results show that spatially clustered missing data degrade performance far more severely than random sparsity, and that link flow data are indispensable for car demand estimation, whereas travel time data strongly influence transit estimation due to intermodal congestion effects. The framework remains relatively stable under varying levels of adaptive routing but is sensitive to the quality of the initial O–D demand. And the matching of passenger count data is more sensitive to behavioral uncertainty compared to other data types. These findings highlight the importance of multi-modal data integration, spatial data diversity, and informed initialization for reliable DODE implementation in real-world networks.