Federated Graph Neural Networks With Equivalent Hypergraph Construction for Traffic Flow Prediction

Traffic flow prediction is essential for intelligent transportation systems, yet privacy concerns and limited cross-regional data sharing hinder accurate modeling of global traffic patterns. This paper proposes a Federated Graph Neural Network with Equivalent Hypergraph (FGNNEH) framework to address these challenges by preserving privacy and enhancing cross-client collaboration. FGNNEH consists of two key stages. First, local traffic networks are transformed into high-dimensional hypernodes through an integrated process of backbone network extraction, kernel matrix analysis, and multilayer perceptrons. The backbone network extraction simplifies graph structures by isolating critical nodes and edges based on topological centrality, ensuring computational efficiency while retaining key spatial dependencies. Kernel matrix analysis captures complex nonlinear correlations among traffic flow features, including spatial-temporal dependencies and region-specific dynamics, enabling more effective feature representation. The multilayer perceptrons further fuse these features into robust hypernode embeddings that encapsulate both structural and traffic flow characteristics. Second, a global hypergraph construction mechanism is introduced to optimize inter-client collaboration. This mechanism employs an iterative performance feedback loop to dynamically add or remove edges between hypernodes, addressing the issue of lost inter-client connections and enabling effective cross-regional information exchange. Together, these components reconstruct a global traffic model that balances local privacy with holistic accuracy. Experiments on real-world traffic datasets, including PeMSD4, METR-LA and Guangzhou, demonstrate that FGNNEH outperforms existing methods in prediction accuracy, computational efficiency, and scalability.