Deep Heterogeneity Learning for Cross-City Transit Forecasting: A Differentially Private Federated Framework with Mixture-of-Experts and Seasonal Decomposition

Aivar Sakhipov^1*Zhanbai Uzdenbayev²Diar Begisbayev^1*Aruzhan Mektepbayeva¹Ramazan Seiitbek¹Didar Yedilkhan¹

• ¹Astana IT University, Astana, Kazakhstan
• ²Zhetysu University named after Ilyas Zhansugurov, Taldykorgan, Kazakhstan

The accurate prediction of transit flows is fundamental to optimizing intelligent transportation systems; however, the deployment of centralized forecasting models is frequently obstructed by the heterogeneous, Non-Independent and Identically Distributed (Non-IID) nature of cross-city data and stringent data privacy regulations. To address these challenges, this study proposes X-FedFormer, a novel framework that integrates Federated Learning (FL) with Differential Privacy (DP) to enable collaborative, privacy-preserving forecasting across geographically disparate cities. The proposed architecture synergistically combines a Mixture-of-Experts (MoE) mechanism to dynamically adapt to diverse local data distributions and a Seasonal-Trend Decomposition module to explicitly disentangle complex, multi-scale temporal dependencies. To overcome the scarcity of open multi-city transit data, the framework is evaluated on a statistically validated synthetic dataset that rigorously simulates realistic inflow and outflow patterns across ten diverse urban environments. Experimental results demonstrate that X-FedFormer significantly outperforms state-of-the-art federated baselines, including FedProx, achieving an aggregate coefficient of determination of 0.922 and a consistent Mean Absolute Error (MAE) profile across all participating cities. Furthermore, ablation studies confirm that the integration of MoE and seasonal decomposition reduces forecasting error by approximately 11% and 16%, respectively, compared to standard architectures. The study also characterizes the privacy-utility frontier, establishing that the model maintains high predictive utility even under strict differential privacy guarantees. These findings present a scalable, robust solution for urban computing, effectively balancing the trade-off between algorithmic performance and data sovereignty in smart city applications.