Communication-Efficient Personalized Federated Learning for Traffic Flow Prediction

As an important component of intelligent transportation systems, accurate traffic flow prediction can improve road traffic efficiency. Traffic data is typically represented as graphs, but data from different regions is independently collected by various organizations (clients), resulting in the division of the complete graph structure into multiple subgraphs. Feder-ated learning enables the collaborative use of these independent datasets. However, limited communication bandwidth and computational resources result in high overhead in practical scenarios, hindering the application of federated learning. To address this, we introduce an efficient pruning-based federated learning algorithm (DPruneFL), which analyzes client differences and retains personalized model parts. Additionally, we developed the FedNe-DP system by integrating this with the personalized federated learning algorithm based on neighbor aggregation (FedNe), considering both the physical relationships among subgraph data and client personalization needs during aggregation. Experimental results on real datasets show that our approach not only improves prediction performance but also reduces communication costs, enhancing the efficiency of federated learning in actual traffic flow prediction.