Sakhipov A., Uzdenbayev Z., Begisbayev D., Mektepbayeva A., Seiitbek R., Yedilkhan D.
Deep Heterogeneity Learning for Cross-City Transit Forecasting: A Differentially Private Federated Framework with Mixture-of-Experts and Seasonal Decomposition // Frontiers in Future Transportation. — 2026. — Vol. 7. — Article No. 1644979. — DOI: 10.3389/ffutr.2026.1644979.

Abstract: Accurate prediction of transit flows is essential for optimizing intelligent transportation systems, but centralized forecasting is often limited by heterogeneous cross-city data and privacy constraints. This paper proposes X-FedFormer, a federated learning framework with differential privacy that integrates a mixture-of-experts mechanism and seasonal-trend decomposition for robust cross-city transit forecasting. The model is evaluated on a synthetic dataset simulating realistic inflow and outflow patterns across ten urban environments, containing 90 days of hourly records per city. Experimental results show that X-FedFormer outperforms federated baselines such as FedProx, achieving a coefficient of determination of 0.922 and a mean absolute error of 7.93 passengers. Statistical validation using the Wilcoxon signed-rank test confirms significant improvements (p = 0.018). Ablation studies demonstrate that mixture-of-experts and seasonal decomposition reduce forecasting error by approximately 11% and 16%, respectively. The framework maintains strong performance under differential privacy constraints (ε ≈ 2), offering a scalable and privacy-preserving solution for smart city transit analytics.

Link / DOI: https://doi.org/10.3389/ffutr.2026.1644979