FedFS-UTP-Dyn: Adaptive Vertical Federated Learning for Urban Traffic Prediction under Concept Drift with Warm Restart, CUSUM Gating, and Federated Gini

Messam Raza; Kinza Tahir

doi:10.63056/academia.5.3(s11).2026.2516

Authors

Messam Raza Department of Computer Science, FAST-NUCES, Islamabad, Pakistan Author
Kinza Tahir Department of Computer Science & IT, Mirpur University of Science & Technology, Mirpur, AJK, Pakistan Author

DOI:

https://doi.org/10.63056/academia.5.3(s11).2026.2516

Keywords:

Vertical federated learning, urban traffic prediction, concept drift, feature selection, CUSUM, warm restart, communication efficiency, Adaptive Feature Selection, CUSUM Drift Detection, Federated Gini Protocol, Vertical Federated Learning, Urban Traffic Prediction, Warm-Restart Layer Rebuild, Concept Drift Recovery, Homomorphic Encryption Elimination

Abstract

Vertical federated learning (VFL) for urban traffic prediction (UTP) requires adaptive feature selection to handle non-stationary data distributions. Existing Gini-based reselection approaches typically reinitialise client input layers after each mask update, disrupting learned representations and introducing performance instability. This paper presents FedFS-UTP-Dyn, a drift-aware VFL framework that integrates three mechanisms: a

warm-restart layer rebuild to preserve weights for retained features, a CUSUM (Page–Hinkley)-based drift detector to trigger feature re-evaluation under sustained performance degradation, and a federated Gini protocol that performs local feature scoring and transmits only binary mask proposals. The proposed framework is evaluated on two real-world traffic datasets and a controlled concept-drift scenario across multiple client configurations and five independent seeds. On the Accident Severity dataset, the warm-restart variant shows a modest improvement over periodic Gini reselection (e.g., 0.8755 vs. 0.8637 AUC at 4C1S, Cohen’s d =0.50), though differences remain statistically non-significant. On the near-stationary PeMS dataset, all methods converge within 0.003 AUC (0.927–0.930), indicating no performance degradation under stable conditions. The CUSUM-based trigger reduces re-evaluation frequency by approximately 79% relative to fixed-interval scheduling, while the federated Gini protocol achieves a 65,536× reduction in communication overhead per reevaluation round. In a controlled drift setting, the adaptive model recovers AUC from 0.497 to 0.836 after abrupt feature-relevance inversion, demonstrating effective adaptation to changing feature importance. These results indicate that FedFS-UTP-Dyn provides an efficient and adaptive approach for feature management in VFL-based traffic prediction systems. Source code is publicly available at our GitHub repository.