FastFlow: Availability Aware Federated Attention-Based Spatial-Temporal GNN for Big Data-Driven Traffic Forecasting

The rapid evolution of Intelligent Transportation Systems (ITS) in the Big Data era, propelled by the Internet of Things (IoT), has led to advanced data-driven vehicle traffic forecasting. Graph Neural Networks (GNNs), particularly the Attention-Based Spatial-Temporal Graph Neural Networks (AST-GNN), are promising in traffic forecasting but face limitations in integrating Big Data with privacy-preserving Federated Learning (FL) due to unique data topology processing. This paper intro-duces the Availability Aware Federated Attention-based Spatial-Temporal Graph Neural Network (FastFlow), an innovative framework that enhances ASTGNN by integrating Federated Learning across entities and employing Big Data methodologies. FastFlow's distinctiveness lies in its availability-aware approach, aggregating adjacency matrices for global topology and utilizing a novel communication protocol that prioritizes data availability and correlation among organizations. Our evaluation of Caltrans Performance Measurement System (PEMS) data in a simulated setup demonstrates FastFlow's ability to balance predictive accuracy and data security in a multi-organizational context.