This paper introduces VFL-MMBD, a privacy-preserving framework that enables collaborative demand forecasting across multiple bike sharing operators without sharing raw data. It integrates a Split Neural Network architecture with Graph Neural Networks to fuse spatiotemporal features from disparate data sources, addressing the vertical data partitioning problem. A novel Multi-Modal Attention Fusion module learns cross-modal interactions while preserving data privacy through homomorphic encryption.
Key findings
VFL-MMBD achieves a mean absolute percentage error reduction of 15.2% over local-only baselines and 8.7% over horizontal federated learning approaches.
The framework integrates a Split Neural Network with Graph Neural Networks to address vertical data partitioning in multi-operator scenarios.
A novel Multi-Modal Attention Fusion module is introduced to learn cross-modal interactions between bike sharing, public transit, and environmental data.
Limitations & open questions
The paper does not discuss potential scalability issues with an increasing number of operators.
The framework's performance under different data granularities and formats across operators is not fully explored.