Traffic Flow Prediction Using Federated Learning

Today, the rapid increase in urbanization and vehicle density in metropolises has made traffic flow forecasting a crucial need in the context of smart city applications. However, the collection of traffic data on central servers also raises concerns about data privacy and security. In this study, a hybrid GraphConv–TKAN architecture based on federated learning has been developed to model spatial and temporal dependencies in urban traffic data, aiming to protect data privacy. In the proposed hybrid model, Graph Convolutional Network (GraphConv) is used to learn spatial dependencies, while temporal dependencies are modeled with the Temporal Kolmogorov–Arnold Network (TKAN) architecture. In addition to TKAN, Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) architectures are employed to enable a comprehensive comparison; to the best of our knowledge, this study is among the first to systematically compare TKAN, LSTM, and BiLSTM models under both centralized and federated learning. The Temporal Graph Convolutional Network (T-GCN) model, as an alternative spatio-temporal approach that addresses both spatial and temporal dependencies, was evaluated for comparison purposes. In this context, extended variants were created by integrating LSTM, BiLSTM, and TKAN layers instead of the GRU structure used in the original T-GCN architecture, and their performances were compared. Furthermore, basic temporal architectures modeling only temporal dependencies were tested under both centralized and federated learning; in these models, data was distributed equally to clients, providing a fair and balanced comparison environment. In graph-based architectures, during the data distribution phase to clients in the federated learning process, the Louvain Community Detection method was used to divide nodes into communities, and each community was defined as a client. The performance of the proposed models was evaluated on the hourly traffic density dataset from the Istanbul Metropolitan Municipality (IMM) Open Data Portal and the PeMS08 dataset published by Caltrans. Since the Istanbul traffic dataset initially lacked a graph-based structure, it was converted into a graph structure to represent spatial relationships and restructured to be suitable for graph-based spatial modeling; the PeMS08 dataset, on the other hand, was used directly while preserving its existing graph structure. Experimental results show that the proposed GraphConv–TKAN architecture offers robust and stable performance in a federated learning environment and effectively captures spatio-temporal dependencies. The proposed approach appears to achieve performance levels close to centralized learning when operating under federated settings that protect data privacy. These results suggest that combining graph-based spatial modeling with TKAN-based temporal learning can help mitigate the performance degradation frequently observed in federated traffic forecasting.