Enhanced origin–destination matrix estimation for the NYC subway using turnstile and Wi-Fi data integration with graph neural networks

Origin–Destination (OD) matrices are essential for understanding travel behavior and supporting short- and long-term planning in complex transit systems. This study examines the New York City subway network using two complementary data sources: Metropolitan Transportation Authority (MTA) turnstile counts and Subway Wi-Fi based OD observations. Turnstile data provide complete coverage of entries and exits but no OD detail, whereas Wi-Fi data capture OD patterns but are sparse and biased. To address these limitations, we apply a Generative Adversarial Network (GAN) framework to complete sparse Wi-Fi OD flows using MTA ridership estimation data as structural guidance, producing fused OD matrices for model training. We assess source compatibility at station and OD-pair levels and train FuseODNet, a graph-based predictor that learns spatial demand structure and reconstructs four hourly OD matrices from each 4-hour turnstile input block. In the routine evaluation, FuseODNet attains the lowest normalized RMSE among the external comparators. Its template ensemble prior accounts for most of the routine RMSE performance, while the residual component reduces absolute and high-flow error. A weekday-hour average attains the lowest mean absolute error. For the Hurricane Ida case, FuseODNet has the lowest OD-share RMSE point estimate among all evaluated methods against a September 2021 MTA weekday-hour average benchmark. Yankees and synthetic evaluations examine performance under a localized surge and controlled stress scenarios. The enhanced OD forecasts can support applications ranging from service scheduling to disruption management and emergency evacuation planning, demonstrating the value of combining generative data fusion with graph based learning for urban transit systems.