Security-Enhanced Spatiotemporal Ride-Hailing Demand Prediction—Part I: Horizontal Federated Learning

The burgeoning demand for ride-hailing services, coupled with the complexity of traffic patterns, underscores the need for advanced demand prediction models. Traditional approaches often suffer from data silos, which prevent the effective utilization of big data separately owned by multiple ride-hailing platforms. In this study, we propose TSAGE, a novel spatiotemporal deep learning model that combines GraphSAGE for spatial feature extraction and Gated Recurrent Units (GRU) for capturing temporal dependencies. To enable privacy-preserving collaborative learning, horizontal federated learning (HFL) is integrated into TSAGE, so that multiple ride-hailing platforms are allowed to jointly train models without sharing raw data. Moreover, we implement HFL using differential privacy technique to protect data privacy by adding noise to the locally trained model parameters before sharing them with the central server. Through a detailed case study in New York City for Uber and Lyft, we found that prediction accuracy for each ride-hailing platform correlates positively with the volume of ride-hailing demand and negatively with the volatility in demand. Consequently, areas with high demand demonstrate superior performance, and larger platforms with less demand volatility also exhibit better performance. Moreover, compared to scenarios where each platform individually predicts ride-hailing demand, HFL showed significant performance enhancements. Additionally, compared with other HFL methods, the HFL applied with TSAGE can effectively address the challenges posed by the non-independent and identically distributed data across different platforms.