Security-Enhanced Spatiotemporal Ride-Hailing Demand Prediction—Part II: Vertical Federated Learning

In Part I of this study, we proposed a TSAGE-based horizontal federated learning (HFL) paradigm tailored for ride-hailing demand forecasting. While HFL allows ride-hailing platforms to collaborate on demand prediction without sharing data, it only expands the sample space and does not utilize complementary features from different providers. To address this limitation, in Part II, we propose vertical federated learning (VFL), which enables multiple platforms to share different feature sets for one client while preserving data privacy by homomorphic encryption. Furthermore, to guarantee data integrity and trustworthiness, we integrate a consortium blockchain framework that records data transactions and model updates, ensuring transparency and preventing malicious data manipulation. Through a case study, in addition to the two insights obtained in Part I, namely that the prediction accuracy is positively correlated with the ride-hailing demand volume and negatively correlated with the demand volatility, we have also discovered two new insights: compared to scenarios where each platform individually predicts ride-hailing demand, both HFL and VFL showed significant performance enhancements. Notably, HFL is slightly faster than VFL in training and inference, whereas VFL slightly outperforms HFL in the metrics considered in the case study. Interestingly, VFL provides more substantial benefits for ride-hailing platforms with lower demand levels (i.e., Lyft), demonstrating VFL’s capability to offer more equitable benefits across stakeholders, even under asymmetric data distributions. This underscores the enhanced accuracy and fairness properties of VFL compared to HFL.