Reliability-Aware Refinement of High-Resolution Human Mobility Origin-Destination Tables under Sparsity and Suppression

Abstract

High-resolution origin-destination (OD) tables derived from digital mobility traces can preserve fine-scale collective movement structure, but sampling, population expansion, and privacy-motivated low-count suppression create heterogeneous observational support across OD entries and make apparent zeros difficult to interpret. We develop a reliability-aware refinement framework that combines each released OD entry with a time-resolved, data-informed radiation baseline through a closed-form cell-wise Gamma-Poisson update, without smoothing neighboring OD pairs. Applied to a 125 m hourly Tokyo OD product, the method preserves concentrated directional flows while adding lower-intensity, baseline-supported links. The relative contribution of released and model-supported structure varies substantially across OD pairs and departure hours, including among positive-flow entries: released flows are most strongly retained at 7:00, followed by 17:00, whereas the radiation baseline contributes more at 13:00. The inferred characteristic radius broadens the movement range represented by the baseline and exhibits structured variation around major station areas and rail corridors. Tokyo and Fukuoka also show a common increase in data-baseline separation at long distances. Under controlled Poisson thinning, calibrated refinement yields lower realization-averaged mean absolute error than an independently resampled no-refinement table across the tested conditions. This experiment provides internal calibration under imposed degradation, not external validation against true unsuppressed flows. The framework thus provides a transparent way to preserve fine-scale empirical mobility structure while making the cell-wise contribution of structured model support explicit.