A Short-Term Traffic Flow Prediction Method Based on Personalized Lightweight Federated Learning.

Area of Science:

• Urban planning and transportation science
• Artificial intelligence and machine learning

Background:

• Accurate traffic flow prediction is crucial for effective land use and urban expansion planning.
• Existing federated learning methods may not fully accommodate the nuances of traffic flow data or ensure personalization.

Purpose of the Study:

• To introduce a novel personalized lightweight federated learning (PLFL) framework tailored for traffic flow prediction.
• To enhance privacy, personalization, and communication efficiency in collaborative traffic flow modeling.

Main Methods:

• Development of a personalized lightweight federated learning (PLFL) framework.
• Utilized a spatiotemporal fusion graph convolutional network (MGTGCN) as the initial model.
• Incorporated customized client weight allocation and dynamic model pruning (DMP) for enhanced personalization and communication efficiency.

Main Results:

• The PLFL framework achieved favorable traffic flow prediction outcomes, even with missing data from certain clients.
• Demonstrated enhanced communication efficiency in federated learning.
• Preserved individual client characteristics without significant interference.

Conclusions:

• The proposed PLFL framework offers an effective solution for privacy-preserving, personalized, and efficient traffic flow prediction.
• The framework shows robustness in handling data heterogeneity and improving communication overhead in federated learning scenarios for urban planning.