Federated Drift-Aware Graph Neural Forecasting for Real-Time Passenger Flow

Abstract

This paper introduces FedST-GNN, a federated spatio-temporal graph neural network that combines encrypted federated averag- ing with frequency-domain Transformers and ADWIN-triggered meta-learning for privacy-preserving passenger flow forecasting. Evaluation on the Copenhagen-Flow dataset (18.7M events, 312 stops) demonstrates 5% and 7% improvements in MAE and RMSE respectively, over Temporal Fusion Transformer baselines, while maintaining 38ms inference latency on commodity hard- ware. During anomalous events such as a city half-marathon, the adaptive drift detection mechanism reduced peak prediction errors by 41% within a 5MB communication budget per federated round. These results establish that federated graph neural architectures can deliver real-time accuracy for intelligent transportation systems while preserving data privacy.

Related publications

• Deep heterogeneity learning for cross-city transit forecasting: a differentially private federated framework with mixture-of-experts and seasonal decompositionFrontiers in Future Transportation · 2026 · Same author: Diar Begisbayev
• Adaptive Machine Learning Algorithms for Data Processing in Transportation SystemsIEEE International Conference on Smart Information Systems and Technologies · 2025 · Citations: 1 · Same author: Diar Begisbayev
• Investigation of Hybrid Models and Architectures for Real-Time Adaptive Passenger Flow PredictionInternational Conference on Data Science and Information Technology · 2024 · Citations: 1 · Same author: Diar Begisbayev
• DETECTING AND RANKING URBAN BOTTLENECKS FROM PASSIVE SPEED AGGREGATESScientific Journal of Astana IT University · 2026 · Shared topic: Traffic Prediction and Management Techniques